NB-Wav2Vec2-300M-Bokmaal-v2

.gitattributes

@@ -2,13 +2,11 @@
 *.arrow filter=lfs diff=lfs merge=lfs -text
 *.bin filter=lfs diff=lfs merge=lfs -text
 *.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
 *.ftz filter=lfs diff=lfs merge=lfs -text
 *.gz filter=lfs diff=lfs merge=lfs -text
 *.h5 filter=lfs diff=lfs merge=lfs -text
 *.joblib filter=lfs diff=lfs merge=lfs -text
 *.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
 *.model filter=lfs diff=lfs merge=lfs -text
 *.msgpack filter=lfs diff=lfs merge=lfs -text
 *.npy filter=lfs diff=lfs merge=lfs -text
@@ -22,14 +20,22 @@
 *.pt filter=lfs diff=lfs merge=lfs -text
 *.pth filter=lfs diff=lfs merge=lfs -text
 *.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
 saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
 *.tflite filter=lfs diff=lfs merge=lfs -text
 *.tgz filter=lfs diff=lfs merge=lfs -text
 *.wasm filter=lfs diff=lfs merge=lfs -text
 *.xz filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
+*.zstandard filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+wandb/run-20220725_150933-2x1p7456 filter=lfs diff=lfs merge=lfs -text
+wandb/debug-internal.log filter=lfs diff=lfs merge=lfs -text
+wandb/debug.log filter=lfs diff=lfs merge=lfs -text
+wandb/latest-run filter=lfs diff=lfs merge=lfs -text
+wandb/run-20220725_150933-2x1p7456/files filter=lfs diff=lfs merge=lfs -text
+wandb/run-20220725_150933-2x1p7456/logs filter=lfs diff=lfs merge=lfs -text
+wandb/run-20220725_150933-2x1p7456/run-2x1p7456.wandb filter=lfs diff=lfs merge=lfs -text
+wandb/run-20220725_150933-2x1p7456/tmp filter=lfs diff=lfs merge=lfs -text
+*.wandb filter=lfs diff=lfs merge=lfs -text
+language_model/unigrams.txt filter=lfs diff=lfs merge=lfs -text

add_kenlm.py

@@ -0,0 +1,36 @@
+import argparse
+from transformers import AutoProcessor
+from transformers import Wav2Vec2ProcessorWithLM
+from pyctcdecode import build_ctcdecoder
+
+
+def main(args):
+    processor = AutoProcessor.from_pretrained(args.model_name_or_path)
+    vocab_dict = processor.tokenizer.get_vocab()
+    sorted_vocab_dict = {
+        k.lower(): v for k, v in sorted(vocab_dict.items(), key=lambda item: item[1])
+    }
+    decoder = build_ctcdecoder(
+        labels=list(sorted_vocab_dict.keys()),
+        kenlm_model_path=args.kenlm_model_path,
+    )
+    processor_with_lm = Wav2Vec2ProcessorWithLM(
+        feature_extractor=processor.feature_extractor,
+        tokenizer=processor.tokenizer,
+        decoder=decoder,
+    )
+    processor_with_lm.save_pretrained(args.model_name_or_path)
+    print(
+        f"Run: ~/bin/build_binary language_model/*.arpa language_model/5gram.bin -T $(pwd) && rm language_model/*.arpa")
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--model_name_or_path', default="./", help='Model name or path. Defaults to ./')
+    parser.add_argument('--kenlm_model_path', required=True, help='Path to KenLM arpa file.')
+    args = parser.parse_args()
+    return args
+
+
+if __name__ == "__main__":
+    main(parse_args())

added_tokens.json

@@ -0,0 +1 @@
+{"<s>": 33, "</s>": 34}

all_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 15.0,
+    "train_loss": 0.17642972585578237,
+    "train_runtime": 792752.3013,
+    "train_samples": 302347,
+    "train_samples_per_second": 5.721,
+    "train_steps_per_second": 0.179
+}

alphabet.json

@@ -0,0 +1 @@
+{"labels": [" ", "a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "q", "r", "s", "t", "u", "v", "w", "x", "y", "z", "\u00e5", "\u00e6", "\u00f8", "\u0125", "\u2047", "", "<s>", "</s>"], "is_bpe": false}

cardinal_numbers.py

@@ -0,0 +1,690 @@
+#!/usr/bin/env python
+# coding: utf-8
+
+"""From https://github.com/peresolb/number-conversion/"""
+import sys
+import os
+import nltk
+
+
+# Dict for basic primitive numbers: 1-10
+b = {
+    1: "én",
+    2: "to",
+    3: "tre",
+    4: "fire",
+    5: "fem",
+    6: "seks",
+    7: "sju",
+    8: "åtte",
+    9: "ni",
+    10: "ti",
+}
+b_nn = {
+    1: "ein",
+    2: "to",
+    3: "tre",
+    4: "fire",
+    5: "fem",
+    6: "seks",
+    7: "sju",
+    8: "åtte",
+    9: "ni",
+    10: "ti",
+}
+
+# Dict for teen primitive numbers: 11-19
+t = {
+    11: "elleve",
+    12: "tolv",
+    13: "tretten",
+    14: "fjorten",
+    15: "femten",
+    16: "seksten",
+    17: "sytten",
+    18: "atten",
+    19: "nitten",
+}
+
+
+# Dict for two digit primitive numbers: 20-90
+do = {
+    20: "tjue",
+    30: "tretti",
+    40: "førti",
+    50: "femti",
+    60: "seksti",
+    70: "sytti",
+    80: "åtti",
+    90: "nitti",
+}
+
+
+# Dict for the 3 digit primitive number: 100
+doo = {100: "hundre"}
+
+
+# Dict for the 4 digit primitive number: 1000
+dooo = {1000: "tusen"}
+
+
+# Reverser function for primitive number dicts
+def _revdict(numberdict):
+    newdict = {}
+    for k, v in numberdict.items():
+        newdict[v] = k
+    return newdict
+
+
+# The reverse of the primitive number dicts, where strings are keys and integers are values,
+# are name of original dict underscore 'r'
+b_r, b_nn_r, t_r, do_r, doo_r, dooo_r = (
+    _revdict(b),
+    _revdict(b_nn),
+    _revdict(t),
+    _revdict(do),
+    _revdict(doo),
+    _revdict(dooo),
+)
+
+
+def _oneten(nr, reverse=False, nn=False):
+    """Function taking an int from 1-10 and returning the corresponding word,
+    or, if reverse=True, taking a numberword from 1-10 and returning the digit"""
+    if reverse == False:
+        if not type(nr) is int:
+            return None
+        if nr <= 10:
+            if nn == False:
+                return b[nr]
+            else:
+                return b_nn[nr]
+    else:
+        if not type(nr) is str:
+            return None
+        if nn == False:
+            if nr in b_r.keys():
+                return b_r[nr]
+        else:
+            if nr in b_nn_r.keys():
+                return b_nn_r[nr]
+
+
+def _onedig(nr, reverse=False, nn=False):
+    if reverse == False:
+        if not _oneten(nr) == "ti":
+            if nn == False:
+                return _oneten(nr)
+            else:
+                return _oneten(nr, nn=True)
+    if reverse == True:
+        if not _oneten(nr, reverse=True) == 10:
+            if nn == False:
+                return _oneten(nr, reverse=True)
+            else:
+                return _oneten(nr, reverse=True, nn=True)
+
+
+def _teen(nr, reverse=False):
+    """Function taking a primitive two-digit int in the teen range and returning the
+    corresponding word, or, if reverse=True, the corresponding number word"""
+    if reverse == False:
+        if not type(nr) is int:
+            return None
+        if nr in t.keys():
+            return t[nr]
+    else:
+        if not type(nr) is str:
+            return None
+        if nr in t_r.keys():
+            return t_r[nr]
+
+
+def _twodig(nr, reverse=False):
+    """Function taking a primitive two-digit int in the range 20-90 and returning
+    the corresponding word, or, if reverse=True, the corresponding number word"""
+    if reverse == False:
+        if not type(nr) is int:
+            return None
+        if nr in do.keys():
+            return do[nr]
+    else:
+        if not type(nr) is str:
+            return None
+        if nr in do_r.keys():
+            return do_r[nr]
+
+
+def _numparser(numword, nn=False):
+    """Parse word to see if they start with a wd in firstnumwords.
+    If yes, return firstnumword and second part"""
+    if not type(numword) is str:
+        return None
+    firstpart = ""
+    scndpart = ""
+    firstnumwords = list(do_r.keys())
+    for s in firstnumwords:
+        if numword.startswith(s):
+            slength = len(s)
+            firstpart = s
+            scndpart = numword[slength:]
+            if nn == False:
+                if (
+                    scndpart in b_r.keys() and b_r[scndpart] < 10
+                ):  # Only return if second part is dig below 10
+                    return (firstpart, scndpart)
+            else:
+                if (
+                    scndpart in b_nn_r.keys() and b_nn_r[scndpart] < 10
+                ):  # Only return if second part is dig below 10
+                    return (firstpart, scndpart)
+
+
+def _one_to_nineteen(nr, reverse=False, nn=False):
+    """Function taking a primitive two-digit int in the range 1-19
+    and returning the corresponding word, or, if reverse=True, the corresponding number word"""
+    if reverse == False:
+        if not type(nr) is int:
+            return None
+        if nr < 11:
+            if nn == False:
+                return _oneten(nr)
+            else:
+                return _oneten(nr, nn=True)
+        elif nr < 20:
+            return _teen(nr)
+    else:
+        if not type(nr) is str:
+            return None
+        if nn == False:
+            if type(_oneten(nr, reverse=True)) is int:
+                return _oneten(nr, reverse=True)
+            elif type(_teen(nr, reverse=True)):
+                return _teen(nr, reverse=True)
+        else:
+            if type(_oneten(nr, reverse=True, nn=True)) is int:
+                return _oneten(nr, reverse=True, nn=True)
+            elif type(_teen(nr, reverse=True)):
+                return _teen(nr, reverse=True)
+
+
+def _one_to_nn(nr, reverse=False, nn=False):
+    """Function taking an int in the range 1-99 and returning the corresponding word. Reverse as before"""
+    if reverse == False:
+        if not type(nr) is int:
+            return None
+        if nr > 0:
+            if nr < 20:
+                if nn == False:
+                    return _one_to_nineteen(nr)
+                else:
+                    return _one_to_nineteen(nr, nn=True)
+            elif nr < 100:
+                if nr in do.keys():
+                    return _twodig(nr)
+                else:
+                    nrstring = str(nr)
+                    frstdig = int(nrstring[0]) * 10
+                    scndig = int(nrstring[1])
+                    frstwd = _twodig(frstdig)
+                    if nn == False:
+                        scnwd = _onedig(scndig)
+                    else:
+                        scnwd = _onedig(scndig, nn=True)
+                    nrwd = frstwd + scnwd
+                    return nrwd
+    else:
+        if not type(nr) is str:
+            return None
+        if nn == False:
+            if type(_one_to_nineteen(nr, reverse=True)) is int:
+                return _one_to_nineteen(nr, reverse=True)
+            elif type(_twodig(nr, reverse=True)) is int:
+                return _twodig(nr, reverse=True)
+            else:
+                if _numparser(nr) == None:
+                    return None
+                parsed = _numparser(nr)
+                first = _twodig(parsed[0], reverse=True)
+                second = _one_to_nineteen(parsed[1], reverse=True)
+                return first + second
+        else:
+            if type(_one_to_nineteen(nr, reverse=True, nn=True)) is int:
+                return _one_to_nineteen(nr, reverse=True, nn=True)
+            elif type(_twodig(nr, reverse=True)) is int:
+                return _twodig(nr, reverse=True)
+            else:
+                if _numparser(nr, nn=True) == None:
+                    return None
+                parsed = _numparser(nr, nn=True)
+                first = _twodig(parsed[0], reverse=True)
+                second = _one_to_nineteen(parsed[1], reverse=True, nn=True)
+                return first + second
+
+
+def _one_to_nnn(nr, reverse=False, nn=False):
+    """Function taking an int in the range 1-999 and returning the corresponding word. Reverse as before"""
+    if reverse == False:
+        if not type(nr) is int:
+            return None
+        if nr == 0:
+            return None
+        if nr < 100:  # 1-99
+            if nn == False:
+                return _one_to_nn(nr)
+            else:
+                return _one_to_nn(nr, nn=True)
+        elif nr < 1000:
+            if nr == 100:  # 100
+                return doo[100]
+            else:
+                nrstring = str(nr)  # 435 181
+                frstdig = int(nrstring[0])  # 4 1
+                scndig = int(nrstring[1])  # 3 8
+                thrdig = int(nrstring[2])  # 5 1
+                scthdig = int(nrstring[1:])  # 35 81
+                if nn == False:
+                    frstwd = _onedig(frstdig)  # fire
+                else:
+                    frstwd = _onedig(frstdig, nn=True)
+                nrwd = ""
+                if scndig == 0:  # 405 or 400
+                    if thrdig == 0:  # 400
+                        nrwd = "%s %s" % (frstwd, doo[100])  # fire hundre
+                    else:  # 405
+                        if nn == False:
+                            thrdwd = _one_to_nn(thrdig)  # fem
+                        else:
+                            thrdwd = _one_to_nn(thrdig, nn=True)
+                        if frstdig != 1:
+                            nrwd = "%s %s og %s" % (
+                                frstwd,
+                                doo[100],
+                                thrdwd,
+                            )  # fire hundre og fem
+                        else:
+                            nrwd = "%s og %s" % (doo[100], thrdwd)  # hundre og fem
+                else:  # 435
+                    scthwd = ""
+                    if nn == False:
+                        scthwd = _one_to_nn(scthdig)  # trettifem
+                    else:
+                        scthwd = _one_to_nn(scthdig, nn=True)
+                    if frstdig != 1:
+                        nrwd = "%s %s og %s" % (frstwd, doo[100], scthwd)
+                    else:
+                        nrwd = "%s og %s" % (doo[100], scthwd)  # hundre og trettifem
+                return nrwd
+    else:
+        if not type(nr) is str:
+            return None
+        if type(doo_r.get(nr, None)) is int:  # hundre - 100
+            return doo_r[nr]
+        elif (
+            len(nr.split(" ")) == 1
+            and type(_one_to_nn(nr, reverse=True)) is int
+            and nn == False
+        ):
+            return _one_to_nn(nr, reverse=True)  # 44
+        elif (
+            len(nr.split(" ")) == 1
+            and type(_one_to_nn(nr, reverse=True, nn=True)) is int
+            and nn == True
+        ):
+            return _one_to_nn(nr, reverse=True, nn=True)  # 44
+        elif len(nr.split(" ")) == 2:  # to hundre
+            splitwords = nr.split(" ")
+            if nn == False and nr == "ett hundre":
+                return 100
+            elif nn == True and nr == "eitt hundre":
+                return 100
+            elif (
+                type(_one_to_nn(splitwords[0], reverse=True)) is int
+                and splitwords[1] == "hundre"
+            ):
+                return _one_to_nn(splitwords[0], reverse=True) * 100
+        elif len(nr.split(" ")) == 3:  # hundre og tre
+            splitwords = nr.split(" ")
+            if splitwords[0] == "hundre" and splitwords[1] == "og":
+                if nn == False:
+                    if type(_one_to_nn(splitwords[2], reverse=True)) is int:
+                        return 100 + _one_to_nn(splitwords[2], reverse=True)
+                else:
+                    if type(_one_to_nn(splitwords[2], reverse=True, nn=True)) is int:
+                        return 100 + _one_to_nn(splitwords[2], reverse=True, nn=True)
+            else:
+                return None
+        elif len(nr.split(" ")) == 4:  # ett hundre og trettifire, fire hundre og åtte
+            splitwords = nr.split(" ")
+            if nn == False:
+                if (
+                    splitwords[0] == "ett"
+                    and splitwords[1] == "hundre"
+                    and splitwords[2] == "og"
+                    and type(_one_to_nn(splitwords[3], reverse=True)) is int
+                ):  # ett hundre og trettifire
+                    return 100 + _one_to_nn(splitwords[3], reverse=True)
+                elif (
+                    type(_one_to_nn(splitwords[0], reverse=True)) is int
+                    and _one_to_nn(splitwords[0], reverse=True) < 10
+                    and splitwords[1] == "hundre"
+                    and splitwords[2] == "og"
+                    and type(_one_to_nn(splitwords[3], reverse=True)) is int
+                ):  # fire hundre og trettifire
+                    hundreds = _one_to_nn(splitwords[0], reverse=True) * 100
+                    tens = _one_to_nn(splitwords[3], reverse=True)
+                    return hundreds + tens
+                else:
+                    return None
+            else:
+                if (
+                    splitwords[0] == "eitt"
+                    and splitwords[1] == "hundre"
+                    and splitwords[2] == "og"
+                    and type(_one_to_nn(splitwords[3], reverse=True, nn=True)) is int
+                ):  # eit hundre og trettifire
+                    return 100 + _one_to_nn(splitwords[3], reverse=True, nn=True)
+                elif (
+                    type(_one_to_nn(splitwords[0], reverse=True, nn=True)) is int
+                    and _one_to_nn(splitwords[0], reverse=True, nn=True) < 10
+                    and splitwords[1] == "hundre"
+                    and splitwords[2] == "og"
+                    and type(_one_to_nn(splitwords[3], reverse=True, nn=True)) is int
+                ):  # fire hundre og trettifire
+                    hundreds = _one_to_nn(splitwords[0], reverse=True, nn=True) * 100
+                    tens = _one_to_nn(splitwords[3], reverse=True, nn=True)
+                    return hundreds + tens
+                else:
+                    return None
+
+
+def _high_hundred(nr, nn=False):
+    """In Norwegian, as in English, it is possible to express the numbers 1100-1999 with hundreds,
+    e.g. "tolv hundre og nittiåtte", /twelve hundred and ninety-eight/. We want to be able to convert
+    these to integers. However, we don't need to produce them, so this algoritm only goes from strings to integers"""
+    if not type(nr) is str:
+        return None
+    if len(nr.split(" ")) > 1:
+        frstwd = nr.split(" ")[0]
+        if not type(_teen(frstwd, reverse=True)) is int:
+            return None
+        frstdig = _teen(frstwd, reverse=True)
+        if len(nr.split(" ")) == 2 and nr.split(" ")[1] == "hundre":  # femten hundre
+            return frstdig * 100
+        elif (
+            len(nr.split(" ")) == 4
+            and nr.split(" ")[1] == "hundre"
+            and nr.split(" ")[2] == "og"
+        ):
+            if (
+                nn == False and type(_one_to_nn(nr.split(" ")[3], reverse=True)) is int
+            ):  # femten hundre og førtito
+                lastdigs = _one_to_nn(nr.split(" ")[3], reverse=True)
+                return (frstdig * 100) + lastdigs
+            elif (
+                nn == True
+                and type(_one_to_nn(nr.split(" ")[3], reverse=True, nn=True)) is int
+            ):  # femten hundre og førtito
+                lastdigs = _one_to_nn(nr.split(" ")[3], reverse=True, nn=True)
+                return (frstdig * 100) + lastdigs
+
+
+def _one_to_nnnnnn(nr, reverse=False, nn=False):
+    """Function taking an int in the range 1-999999 and returning the corresponding word. Reverse as before"""
+    if reverse == False:
+        if not type(nr) is int:
+            return None
+        if nr == 0:
+            return None
+        if nr < 1000:  # 1-999
+            if nn == False:
+                return _one_to_nnn(nr)
+            else:
+                return _one_to_nnn(nr, nn=True)
+        elif nr < 1000000:  # 1000-999999
+            if nr == 1000:  # 1000
+                if nn == False:
+                    return "ett tusen"
+                else:
+                    return "eitt tusen"
+            else:
+                nrstring = str(nr)  # Starting with last three digits. e.g. 23[456]
+                ultdig = int(nrstring[-1])  # 6
+                penultdig = int(nrstring[-2])  # 5
+                antepenultdig = int(nrstring[-3])  # 4
+                ult_and_penultdig = int(nrstring[-2:])  # 56
+                ult_penult_antepenultdig = int(nrstring[-3:])  # 456
+                tailstring = ""
+                if antepenultdig == 0:  # 012, 002, 000
+                    if penultdig == 0:  # 000, 002
+                        if ultdig == 0:  # 000
+                            tailstring = "tusen"
+                        else:  # 002
+                            if nn == False:
+                                ultstring = _one_to_nnn(ultdig)
+                                tailstring = "tusen og %s" % ultstring  # tusen og to
+                            else:
+                                ultstring = _one_to_nnn(ultdig, nn=True)
+                                tailstring = "tusen og %s" % ultstring  # tusen og to
+                    else:  # 012
+                        if nn == False:
+                            ult_and_penultstring = _one_to_nnn(ult_and_penultdig)
+                            tailstring = (
+                                "tusen og %s" % ult_and_penultstring
+                            )  # tusen og tolv
+                        else:
+                            ult_and_penultstring = _one_to_nnn(
+                                ult_and_penultdig, nn=True
+                            )
+                            tailstring = (
+                                "tusen og %s" % ult_and_penultstring
+                            )  # tusen og tolv
+                else:  # 456
+                    if nn == False:
+                        ult_penult_antepenultstring = _one_to_nnn(
+                            ult_penult_antepenultdig
+                        )
+                        if str(ult_penult_antepenultdig)[0] == "1":
+                            tailstring = (
+                                "tusen ett %s" % ult_penult_antepenultstring
+                            )  # tusen ett hundre
+                        else:
+                            tailstring = (
+                                "tusen %s" % ult_penult_antepenultstring
+                            )  # tusen fire hundre og femtiseks
+                    else:
+                        ult_penult_antepenultstring = _one_to_nnn(
+                            ult_penult_antepenultdig, nn=True
+                        )
+                        if str(ult_penult_antepenultdig)[0] == "1":
+                            tailstring = (
+                                "tusen eitt %s" % ult_penult_antepenultstring
+                            )  # tusen ett hundre
+                        else:
+                            tailstring = (
+                                "tusen %s" % ult_penult_antepenultstring
+                            )  # tusen fire hundre og femtiseks
+                startdigs = int(
+                    nrstring[:-3]
+                )  # startstring can consist of the 1, 2 or 3 first digits
+                startstring = ""
+                if startdigs == 1:  # 1001 starts with "ett"
+                    if nn == False:
+                        startstring = "ett"
+                    else:
+                        startstring = "eitt"
+                elif startdigs > 99 and startdigs < 200:  # 155555 starts with ett
+                    if nn == False:
+                        startnumstring = _one_to_nnn(startdigs)
+                        startstring = "ett %s" % startnumstring
+                    else:
+                        startnumstring = _one_to_nnn(startdigs, nn=True)
+                        startstring = "eitt %s" % startnumstring
+                else:  # the remaining numbers are purely compositional
+                    if nn == False:
+                        startstring = _one_to_nnn(startdigs)
+                    else:
+                        startstring = _one_to_nnn(startdigs, nn=True)
+                numstring = "%s %s" % (startstring, tailstring)
+                return numstring
+    else:
+        if not type(nr) is str:
+            return None
+        if type(_one_to_nnn(nr, reverse=True)) is int and nn == False:
+            return _one_to_nnn(nr, reverse=True)  # 444
+        elif type(_one_to_nnn(nr, reverse=True, nn=True)) is int and nn == True:
+            return _one_to_nnn(nr, reverse=True, nn=True)  # 444
+        elif nr == "tusen":  # tusen - 1000
+            return 1000
+        elif (
+            len(nr.split(" ")) > 1 and nr.split(" ")[-1] == "tusen"
+        ):  # ett tusen, ett hundre tusen etc.
+            wdlist = nr.split(" ")
+            firstphrase = " ".join(wdlist[:-1])
+            if type(_one_to_nnn(firstphrase, reverse=True)) is int and nn == False:
+                firstdig = _one_to_nnn(firstphrase, reverse=True)
+                return firstdig * 1000
+            elif (
+                type(_one_to_nnn(firstphrase, reverse=True, nn=True)) is int
+                and nn == True
+            ):
+                firstdig = _one_to_nnn(firstphrase, reverse=True, nn=True)
+                return firstdig * 1000
+            elif (
+                len(nr.split(" ")) == 2 and nr.split(" ")[0] == "ett" and nn == False
+            ):  # ett tusen
+                return 1000
+            elif len(nr.split(" ")) == 2 and nr.split(" ")[0] == "eitt" and nn == True:
+                return 1000
+            else:  # misspellings should not result in return value
+                return None
+        else:
+            if len(nr.split(" ")) > 1:  # all other numbers should contain spaces
+                numwordlist = nr.split(" ")
+                if (
+                    "tusen" in numwordlist
+                ):  # Find last part of numphrase, which starts with "tusen"
+                    tusenind = numwordlist.index("tusen")  # find index of "tusen"
+                    lastwords = numwordlist[tusenind:]  # words from 'tusen'
+                    firstwords = numwordlist[:tusenind]  # words until 'tusen'
+                    lastdigs = 0
+                    if len(lastwords) == 3:
+                        if lastwords[1] == "og":  # 'tusen og fire' 'tusen og førtifire'
+                            lastword = lastwords[-1]
+                            if nn == False:
+                                lastdigs = _one_to_nnn(lastword, reverse=True)
+                            elif nn == True:
+                                lastdigs = _one_to_nnn(lastword, reverse=True, nn=True)
+                        elif (
+                            nn == False
+                            and type(_one_to_nnn(" ".join(lastwords[1:]), reverse=True))
+                            is int
+                            and lastwords[2] == "hundre"
+                        ):  # tusen to hundre
+                            hundredphrase = " ".join(lastwords[1:])
+                            lastdigs = _one_to_nnn(hundredphrase, reverse=True)
+                        elif (
+                            nn == True
+                            and type(
+                                _one_to_nnn(
+                                    " ".join(lastwords[1:]), reverse=True, nn=True
+                                )
+                            )
+                            is int
+                            and lastwords[2] == "hundre"
+                        ):  # tusen to hundre
+                            hundredphrase = " ".join(lastwords[1:])
+                            lastdigs = _one_to_nnn(hundredphrase, reverse=True, nn=True)
+                        else:  # misspellings should not result in return value
+                            return None
+                    elif (
+                        len(lastwords) == 5 and lastwords[2] == "hundre"
+                    ):  # 'tusen fire hundre og fem'
+                        hundredphrase = " ".join(lastwords[1:])
+                        if nn == False:
+                            lastdigs = _one_to_nnn(hundredphrase, reverse=True)
+                        else:
+                            lastdigs = _one_to_nnn(hundredphrase, reverse=True, nn=True)
+                    else:  # misspellings should not result in return value
+                        return None
+                    firstdigs = 0
+                    firstphrase = " ".join(firstwords)
+                    if len(firstwords) == 0:  # as in 'tusen og tretti'
+                        firstdigs = 1000
+                    elif (
+                        len(firstwords) == 1 and firstwords[0] == "ett" and nn == False
+                    ):
+                        firstdigs = 1000
+                    elif (
+                        len(firstwords) == 1 and firstwords[0] == "eitt" and nn == True
+                    ):
+                        firstdigs = 1000
+                    elif (
+                        type(_one_to_nnn(firstphrase, reverse=True)) is int
+                        and nn == False
+                    ):
+                        firstdigs = _one_to_nnn(firstphrase, reverse=True) * 1000
+                    elif (
+                        type(_one_to_nnn(firstphrase, reverse=True, nn=True)) is int
+                        and nn == True
+                    ):
+                        firstdigs = (
+                            _one_to_nnn(firstphrase, reverse=True, nn=True) * 1000
+                        )
+                    else:  # misspellings should not result in return value
+                        return None
+                    if type(firstdigs) is int and type(lastdigs) is int:
+                        return firstdigs + lastdigs
+
+
+def convert_nums(nr, reverse=False, nn=False):
+    """Functions for converting numbers. Only works for numbers in range 0-999999 for now"""
+    if reverse == False:
+        if type(nr) is int:
+            returnstring = ""
+            if nr == 0:
+                returnstring = "null"
+            elif nr < 1000000:
+                if nn == False:
+                    returnstring = _one_to_nnnnnn(nr)
+                else:
+                    returnstring = _one_to_nnnnnn(nr, nn=True)
+            else:
+                return None
+            return returnstring
+    else:
+        if type(nr) is str:
+            returnint = 0
+            if nr == "null":
+                returnint = 0
+            elif nn == False and type(_one_to_nnnnnn(nr, reverse=True)) is int:
+                returnint = _one_to_nnnnnn(nr, reverse=True)
+            elif nn == True and type(_one_to_nnnnnn(nr, reverse=True, nn=True)) is int:
+                returnint = _one_to_nnnnnn(nr, reverse=True, nn=True)
+            elif nn == False and type(_high_hundred(nr)) is int:
+                returnint = _high_hundred(nr)
+            elif nn == True and type(_high_hundred(nr, nn=True)) is int:
+                returnint = _high_hundred(nr, nn=True)
+            else:
+                return None
+            return returnint
+
+
+if __name__ == "__main__":
+    # testing
+    mydigit = 243564
+    mynumstring = "hundre og atten tusen fire hundre og trettién"
+    mydigit_nn = 34381
+    mynumstring_nn = "tre hundre og førtiein"
+
+    print(
+        "Digit conversion bm: %s\nString conversion bm: %s"
+        % (convert_nums(mydigit), convert_nums(mynumstring, reverse=True))
+    )
+    print(
+        "Digit conversion nn: %s\nString conversion nn: %s"
+        % (
+            convert_nums(mydigit_nn, nn=True),
+            convert_nums(mynumstring_nn, nn=True, reverse=True),
+        )
+    )

config.json

@@ -0,0 +1,115 @@
+{
+  "_name_or_path": "KBLab/wav2vec2-large-voxrex",
+  "activation_dropout": 0.055,
+  "adapter_kernel_size": 3,
+  "adapter_stride": 2,
+  "add_adapter": false,
+  "apply_spec_augment": true,
+  "architectures": [
+    "Wav2Vec2ForCTC"
+  ],
+  "attention_dropout": 0.094,
+  "bos_token_id": 1,
+  "classifier_proj_size": 256,
+  "codevector_dim": 768,
+  "contrastive_logits_temperature": 0.1,
+  "conv_bias": true,
+  "conv_dim": [
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512
+  ],
+  "conv_kernel": [
+    10,
+    3,
+    3,
+    3,
+    3,
+    2,
+    2
+  ],
+  "conv_stride": [
+    5,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2
+  ],
+  "ctc_loss_reduction": "mean",
+  "ctc_zero_infinity": true,
+  "diversity_loss_weight": 0.1,
+  "do_stable_layer_norm": true,
+  "eos_token_id": 2,
+  "feat_extract_activation": "gelu",
+  "feat_extract_dropout": 0.0,
+  "feat_extract_norm": "layer",
+  "feat_proj_dropout": 0.04,
+  "feat_quantizer_dropout": 0.0,
+  "final_dropout": 0.0,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.047,
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-05,
+  "layerdrop": 0.041,
+  "mask_channel_length": 10,
+  "mask_channel_min_space": 1,
+  "mask_channel_other": 0.0,
+  "mask_channel_prob": 0.0,
+  "mask_channel_selection": "static",
+  "mask_feature_length": 64,
+  "mask_feature_min_masks": 0,
+  "mask_feature_prob": 0.25,
+  "mask_time_length": 10,
+  "mask_time_min_masks": 2,
+  "mask_time_min_space": 1,
+  "mask_time_other": 0.0,
+  "mask_time_prob": 0.082,
+  "mask_time_selection": "static",
+  "model_type": "wav2vec2",
+  "num_adapter_layers": 3,
+  "num_attention_heads": 16,
+  "num_codevector_groups": 2,
+  "num_codevectors_per_group": 320,
+  "num_conv_pos_embedding_groups": 16,
+  "num_conv_pos_embeddings": 128,
+  "num_feat_extract_layers": 7,
+  "num_hidden_layers": 24,
+  "num_negatives": 100,
+  "output_hidden_size": 1024,
+  "pad_token_id": 32,
+  "proj_codevector_dim": 768,
+  "tdnn_dilation": [
+    1,
+    2,
+    3,
+    1,
+    1
+  ],
+  "tdnn_dim": [
+    512,
+    512,
+    512,
+    512,
+    1500
+  ],
+  "tdnn_kernel": [
+    5,
+    3,
+    3,
+    1,
+    1
+  ],
+  "torch_dtype": "float32",
+  "transformers_version": "4.18.0",
+  "use_weighted_layer_sum": false,
+  "vocab_size": 35,
+  "xvector_output_dim": 512
+}

eval.py

@@ -0,0 +1,255 @@
+#!/usr/bin/env python3
+import argparse
+import re
+from typing import Dict
+
+import torch
+from datasets import Audio, Dataset, load_dataset, load_metric
+from num2words import num2words as n2w
+from slugify import slugify
+
+from transformers import AutoFeatureExtractor, AutoModelForCTC, pipeline, Wav2Vec2Processor, Wav2Vec2ProcessorWithLM, Wav2Vec2FeatureExtractor
+# from pyctcdecode import BeamSearchDecoderCTC
+
+from cardinal_numbers import convert_nums
+
+
+def log_results(result: Dataset, args: Dict[str, str]):
+    """DO NOT CHANGE. This function computes and logs the result metrics."""
+
+    log_outputs = args.log_outputs
+    lm = "withLM" if args.use_lm else "noLM"
+    model_id = args.model_id.replace("/", "_").replace(".", "")
+    if args.filter:
+        extra_args = [args.config, slugify(args.filter), args.split, lm]
+    else:
+        extra_args = [args.config, args.split, lm]
+    dataset_id = "_".join([model_id] + args.dataset.split("/") + extra_args)
+
+    # load metric
+    wer = load_metric("wer")
+    cer = load_metric("cer")
+
+    # compute metrics
+    wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
+    cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
+
+    # print & log results
+    result_str = f"{dataset_id}\nWER: {wer_result}\nCER: {cer_result}"
+    print(result_str)
+
+    with open(f"{dataset_id}_eval_results.txt", "w") as f:
+        f.write(result_str)
+    with open(f"{dataset_id}_eval_results.tsv", "w") as f:
+        f.write("\t".join([args.model_id, args.dataset, args.config, args.filter, args.split, str(lm), str(wer_result), str(cer_result)]))
+
+    # log all results in text file. Possibly interesting for analysis
+    if log_outputs is not None:
+        pred_file = f"log_{dataset_id}_predictions.txt"
+        target_file = f"log_{dataset_id}_targets.txt"
+
+        with open(pred_file, "w") as p, open(target_file, "w") as t:
+            # mapping function to write output
+            def write_to_file(batch, i):
+                p.write(f"{i}" + "\n")
+                p.write(batch["prediction"] + "\n")
+                t.write(f"{i}" + "\n")
+                t.write(batch["target"] + "\n")
+
+            result.map(write_to_file, with_indices=True)
+
+
+def normalize_text(original_text: str, dataset: str) -> str:
+    """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
+
+    text = original_text.lower()
+    if dataset.lower().endswith("fleurs"): 
+        replacements = (
+            (r"\be\.kr", "etter kristus fødsel"),
+            (r"\bf\.kr", "før kristi fødsel"),
+            (r"\bca[.]?\b", "circa"),
+            (r"(\d)\s*km/t", r"\1 kilometer i timen"),
+            (r"(\d)\s*km", r"\1 kilometer"),
+            (r"(\d)\s*cm", r"\1 centimeter"),
+            (r"(\d)\s*mm", r"\1 millimeter"),
+            (r"kl\.", "klokka"),
+            (r"f\.eks", "for eksempel"),
+        )
+        for abrev, expasion in replacements:
+            text = re.sub(abrev, expasion, text)
+        text = re.sub(r'(\d+)[-–](\d+)', r'\1 til \2', text)  # 1-89, 70-90
+        text = re.sub(r'(\d{2}):00', r'\1', text)  # 21:00
+        text = re.sub(r"(\d{2}):0(\d{1})", r"\1 null \2", text)  # 17:03
+        text = re.sub(r"(\d{1,2}):(\d{1,2})", r"\1 \2", text)  # 17:23 (time), 4:3 (aspect ratios)
+        text = re.sub(r"(1[1-9])00", r"\1 hundre", text)  # 1800, 1900
+        text = re.sub(r"(1[1-9])0([1-9])", r"\1 null \2 ", text)  # 1901, 1909
+        text = re.sub(r"(1[1-9])([1-9]\d)", r"\1 \2 ", text)  # 1911, 1987
+        text = re.sub(r"(20)0([1-9])", r"\1 null \2 ", text)  # 2009
+        text = re.sub(r"(20)(\d{2})", r"\1 \2 ", text)  # 2009
+        text = re.sub(r"(\d{1,3})[.](\d{1,2})", r"\1 dot \2 ", text)  # 802.11n, 2.5ghz (in English)
+        text = re.sub(r"(\d{1,2})[ .](\d{3})", r"\1\2", text)  # 10 000, 32.000
+        text = re.sub(r'(\w+)-(\w+)', r'\1 \2', text)  # n-standard
+        # text = re.compile(r"-?0?[1-9][\d.]*").sub(lambda x: n2w(x.group(0), lang="no"), text.replace(".", ""))
+        text = re.compile(r"-?0?[1-9][\d.]*").sub(lambda x: convert_nums(int(x.group(0)), nn=True), text.replace(".", ""))
+
+
+    chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\–\_\\\+\#\/]'  # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
+    text = re.sub(chars_to_ignore_regex, "", text) + " "
+
+    if dataset.lower().endswith("nst"):
+        text = text.lower()
+        text = text.replace("(...vær stille under dette opptaket...)", "")
+        text = re.sub('[áàâ]', 'a', text)
+        text = re.sub('[ä]', 'æ', text)
+        text = re.sub('[éèëê]', 'e', text)
+        text = re.sub('[íìïî]', 'i', text)
+        text = re.sub('[óòöô]', 'o', text)
+        text = re.sub('[ö]', 'ø', text)
+        text = re.sub('[ç]', 'c', text)
+        text = re.sub('[úùüû]', 'u', text)
+        # text = re.sub('\\(?=(Punktum|Komma|Utropstegn|Spørsmålstegn))', ' ', text)
+        text = re.sub('\s+', ' ', text)
+    elif dataset.lower().endswith("npsc"):
+        text = re.sub('[áàâ]', 'a', text)
+        text = re.sub('[ä]', 'æ', text)
+        text = re.sub('[éèëê]', 'e', text)
+        text = re.sub('[íìïî]', 'i', text)
+        text = re.sub('[óòöô]', 'o', text)
+        text = re.sub('[ö]', 'ø', text)
+        text = re.sub('[ç]', 'c', text)
+        text = re.sub('[úùüû]', 'u', text)
+        text = re.sub('\s+', ' ', text)
+    elif dataset.lower().endswith("fleurs"): 
+        text = re.sub('[áàâ]', 'a', text)
+        text = re.sub('[ä]', 'æ', text)
+        text = re.sub('[éèëê]', 'e', text)
+        text = re.sub('[íìïî]', 'i', text)
+        text = re.sub('[óòöô]', 'o', text)
+        text = re.sub('[ö]', 'ø', text)
+        text = re.sub('[ç]', 'c', text)
+        text = re.sub('[úùüû]', 'u', text)
+        text = re.sub('[«»]', '', text)
+        text = re.sub('\s+', ' ', text)
+    text = re.sub('<e+h?>', 'ĥ', text)
+    text = re.sub('<m+>', 'ĥ', text)
+    text = re.sub('<q+>', 'ĥ', text)
+    text = re.sub('<inaudible>', 'ĥ', text)
+    text = re.sub('[<>]', '', text)
+
+    # # In addition, we can normalize the target text, e.g. removing new lines characters etc...
+    # # note that order is important here!
+    # token_sequences_to_ignore = ["\n\n", "\n", "   ", "  "]
+
+    # for t in token_sequences_to_ignore:
+    #     text = " ".join(text.split(t))
+
+    return text.strip()
+
+
+def main(args):
+    # load dataset
+    dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
+    if args.filter:
+        attribute, value = list(map(str.strip, args.filter.split(":")))
+        dataset = dataset.filter(
+            lambda x: x[attribute] == value,
+            desc=f"Filtering on {args.filter}",
+        )
+    # for testing: only process the first two examples as a test
+    # dataset = dataset.select(range(10))
+
+    # load processor
+    feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
+    sampling_rate = feature_extractor.sampling_rate
+
+    # resample audio
+    dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
+
+    # load eval pipeline
+    if args.device is None:
+        args.device = 0 if torch.cuda.is_available() else -1
+    # asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
+
+    model_instance = AutoModelForCTC.from_pretrained(args.model_id)
+    if args.use_lm:
+        processor = Wav2Vec2ProcessorWithLM.from_pretrained(args.model_id)
+        decoder = processor.decoder
+    else:
+        processor = Wav2Vec2Processor.from_pretrained(args.model_id)
+        decoder = None
+    asr = pipeline(
+        "automatic-speech-recognition",
+        model=model_instance,
+        tokenizer=processor.tokenizer, 
+        feature_extractor=processor.feature_extractor,
+        decoder=decoder,
+        device=args.device
+    )
+
+    # feature_extractor_dict, _ = Wav2Vec2FeatureExtractor.get_feature_extractor_dict(args.model_id)
+    # feature_extractor_dict["processor_class"] = "Wav2Vec2Processor" if not args.use_lm else "Wav2Vec2ProcessorWithLM"
+    # feature_extractor = Wav2Vec2FeatureExtractor.from_dict(feature_extractor_dict)
+
+    # asr = pipeline("automatic-speech-recognition", model=args.model_id, feature_extractor=feature_extractor, device=args.device, decoder=BeamSearchDecoderCTC.load_from_dir("./"))
+
+    # map function to decode audio
+    def map_to_pred(batch):
+        prediction = asr(
+            batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
+        )
+
+        batch["prediction"] = prediction["text"]
+        batch["target"] = normalize_text(batch[args.text_column], args.dataset)
+        return batch
+
+    # run inference on all examples
+    result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
+
+    # compute and log_results
+    # do not change function below
+    log_results(result, args)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
+    )
+    parser.add_argument(
+        "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'`  for Common Voice"
+    )
+    parser.add_argument(
+        "--filter", type=str, default="", help="Simple filter on attributes. *E.g.* `region_of_youth:Troms` would pnly keep those samplesfor which the condition is met"
+    )
+    parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
+    parser.add_argument(
+        "--text_column", type=str, default="text", help="Column name containing the transcription."
+    )
+    parser.add_argument(
+        "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
+    )
+    parser.add_argument(
+        "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
+    )
+    parser.add_argument(
+        "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
+    )
+    parser.add_argument(
+        "--device",
+        type=int,
+        default=None,
+        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
+    )
+    parser.add_argument(
+        "--use_lm", action="store_true", help="If defined, use included language model as the decoder."
+    )
+    args = parser.parse_args()
+
+    main(args)

language_model/5gram.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b41c24c63f2f0585bea83666369593f3b3e6d047f327a90f36ebca2c35ef0ff
+size 4243671427

language_model/attrs.json

@@ -0,0 +1 @@
+{"alpha": 0.5, "beta": 1.5, "unk_score_offset": -10.0, "score_boundary": true}

language_model/unigrams.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac3e71ca49838ca355df6fdcb8d89344a5a9bf9e1a76587cdf5df1367c19b9a9
+size 16759269

preprocessor_config.json

@@ -0,0 +1,10 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0,
+  "processor_class": "Wav2Vec2ProcessorWithLM",
+  "return_attention_mask": true,
+  "sampling_rate": 16000
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ecb597e9e18695f8f63e8b845ec3c635e8e559f7da920072736f443e26ca568
+size 1262042225

run.sh

@@ -0,0 +1,39 @@
+WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_speech_recognition_ctc.py \
+        --model_name_or_path="KBLab/wav2vec2-large-voxrex" \
+        --hub_model_id="NbAiLab/wav2vec2-large-voxrex-npsc-nst-bokmaal-fixed" \
+        --output_dir="./" \
+        --overwrite_output_dir \
+        --num_train_epochs="15" \
+        --per_device_train_batch_size="16" \
+        --per_device_eval_batch_size="16" \
+        --gradient_accumulation_steps="2" \
+        --learning_rate="1e-4" \
+        --warmup_steps="2000" \
+        --length_column_name="input_length" \
+        --evaluation_strategy="steps" \
+        --text_column_name="text" \
+        --save_steps="500" \
+        --eval_steps="500" \
+        --logging_steps="100" \
+        --layerdrop="0.041" \
+        --attention_dropout="0.094" \
+        --activation_dropout="0.055" \
+        --hidden_dropout="0.047" \
+        --save_total_limit="3" \
+        --freeze_feature_encoder \
+        --feat_proj_dropout="0.04" \
+        --mask_time_prob="0.082" \
+        --mask_time_length="10" \
+        --mask_feature_prob="0.25" \
+        --mask_feature_length="64" \
+        --gradient_checkpointing \
+        --min_duration_in_seconds="0.5" \
+        --max_duration_in_seconds="20.0" \
+        --use_auth_token \
+        --seed="42" \
+        --fp16 \
+        --group_by_length \
+        --do_train --do_eval \
+        --push_to_hub \
+        --preprocessing_num_workers="32" \
+        --ctc_zero_infinity

run_speech_recognition_ctc.py

@@ -0,0 +1,822 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+
+""" Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
+
+import functools
+import json
+import logging
+import os
+import re
+import sys
+import warnings
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Union
+
+import datasets
+import numpy as np
+import torch
+from datasets import DatasetDict, load_dataset, load_metric
+
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoModelForCTC,
+    AutoProcessor,
+    AutoTokenizer,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    Wav2Vec2Processor,
+    set_seed,
+)
+from transformers.trainer_utils import get_last_checkpoint, is_main_process
+from transformers.utils import check_min_version
+from transformers.utils.versions import require_version
+
+# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
+check_min_version("4.16.0.dev0")
+
+require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
+
+logger = logging.getLogger(__name__)
+
+
+def list_field(default=None, metadata=None):
+    return field(default_factory=lambda: default, metadata=metadata)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    tokenizer_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
+    )
+    freeze_feature_encoder: bool = field(
+        default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
+    )
+    attention_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
+    )
+    activation_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
+    )
+    feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
+    hidden_dropout: float = field(
+        default=0.0,
+        metadata={
+            "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
+        },
+    )
+    final_dropout: float = field(
+        default=0.0,
+        metadata={"help": "The dropout probability for the final projection layer."},
+    )
+    mask_time_prob: float = field(
+        default=0.05,
+        metadata={
+            "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
+                    "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
+                    "vectors will be masked along the time axis."
+        },
+    )
+    mask_time_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the time axis."},
+    )
+    mask_feature_prob: float = field(
+        default=0.0,
+        metadata={
+            "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
+                    "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
+        },
+    )
+    mask_feature_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the feature axis."},
+    )
+    layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
+    ctc_loss_reduction: Optional[str] = field(
+        default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
+    )
+    ctc_zero_infinity: Optional[bool] = field(
+        default=False, metadata={"help": "If True, will try yo aboud the CTC loss goinf to infinity."}
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+
+    Using `HfArgumentParser` we can turn this class
+    into argparse arguments to be able to specify them on
+    the command line.
+    """
+
+    # dataset_name: str = field(
+    #     metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    # )
+    # dataset_config_name: str = field(
+    #     default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    # )
+    train_split_name: str = field(
+        default="train",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    eval_split_name: str = field(
+        default="test",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    audio_column_name: str = field(
+        default="audio",
+        metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
+    )
+    text_column_name: str = field(
+        default="text",
+        metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+                    "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
+                    "value if set."
+        },
+    )
+    chars_to_ignore: Optional[List[str]] = list_field(
+        default=None,
+        metadata={"help": "A list of characters to remove from the transcripts."},
+    )
+    eval_metrics: List[str] = list_field(
+        default=["wer"],
+        metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
+    )
+    max_duration_in_seconds: float = field(
+        default=20.0,
+        metadata={
+            "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
+    )
+    preprocessing_only: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to only do data preprocessing and skip training. "
+                    "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
+                    "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
+                    "so that the cached datasets can consequently be loaded in distributed training"
+        },
+    )
+    use_auth_token: bool = field(
+        default=False,
+        metadata={
+            "help": "If :obj:`True`, will use the token generated when running"
+                    ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
+        },
+    )
+    unk_token: str = field(
+        default="[UNK]",
+        metadata={"help": "The unk token for the tokenizer"},
+    )
+    pad_token: str = field(
+        default="[PAD]",
+        metadata={"help": "The padding token for the tokenizer"},
+    )
+    word_delimiter_token: str = field(
+        default="|",
+        metadata={"help": "The word delimiter token for the tokenizer"},
+    )
+    phoneme_language: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The target language that should be used be"
+                    " passed to the tokenizer for tokenization. Note that"
+                    " this is only relevant if the model classifies the"
+                    " input audio to a sequence of phoneme sequences."
+        },
+    )
+
+
+@dataclass
+class DataCollatorCTCWithPadding:
+    """
+    Data collator that will dynamically pad the inputs received.
+    Args:
+        processor (:class:`~transformers.AutoProcessor`)
+            The processor used for proccessing the data.
+        padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
+            Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
+            among:
+            * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+              sequence if provided).
+            * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
+              maximum acceptable input length for the model if that argument is not provided.
+            * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
+              different lengths).
+        max_length (:obj:`int`, `optional`):
+            Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
+        max_length_labels (:obj:`int`, `optional`):
+            Maximum length of the ``labels`` returned list and optionally padding length (see above).
+        pad_to_multiple_of (:obj:`int`, `optional`):
+            If set will pad the sequence to a multiple of the provided value.
+            This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
+            7.5 (Volta).
+    """
+
+    processor: AutoProcessor
+    padding: Union[bool, str] = "longest"
+    pad_to_multiple_of: Optional[int] = None
+    pad_to_multiple_of_labels: Optional[int] = None
+
+    def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
+        # split inputs and labels since they have to be of different lenghts and need
+        # different padding methods
+        input_features = [{"input_values": feature["input_values"]} for feature in features]
+        label_features = [{"input_ids": feature["labels"]} for feature in features]
+
+        batch = self.processor.pad(
+            input_features,
+            padding=self.padding,
+            pad_to_multiple_of=self.pad_to_multiple_of,
+            return_tensors="pt",
+        )
+
+        with self.processor.as_target_processor():
+            labels_batch = self.processor.pad(
+                label_features,
+                padding=self.padding,
+                pad_to_multiple_of=self.pad_to_multiple_of_labels,
+                return_tensors="pt",
+            )
+
+        # replace padding with -100 to ignore loss correctly
+        labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
+
+        batch["labels"] = labels
+
+        return batch
+
+
+def create_vocabulary_from_data(
+        datasets: DatasetDict,
+        word_delimiter_token: Optional[str] = None,
+        unk_token: Optional[str] = None,
+        pad_token: Optional[str] = None,
+):
+    # Given training and test labels create vocabulary
+    alphabet = set()
+
+    def extract_all_chars(batch):
+        all_text = " ".join(batch["target_text"])
+        alphabet.update(all_text)
+
+    datasets.map(
+        extract_all_chars,
+        batched=True,
+        batch_size=-1,
+        keep_in_memory=True,
+        remove_columns=datasets["train"].column_names,
+    )
+
+    # # take union of all unique characters in each dataset
+    # vocab_set = functools.reduce(
+    #     lambda vocab_1, vocab_2: {"vocab": list(set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]))}, vocabs.values()
+    # )["vocab"][0]
+
+    vocab_dict = {v: k for k, v in enumerate(sorted(list(alphabet)))}
+
+    # replace white space with delimiter token
+    if word_delimiter_token is not None:
+        vocab_dict[word_delimiter_token] = vocab_dict[" "]
+        del vocab_dict[" "]
+
+    # add unk and pad token
+    if unk_token is not None:
+        vocab_dict[unk_token] = len(vocab_dict)
+
+    if pad_token is not None:
+        vocab_dict[pad_token] = len(vocab_dict)
+
+    return vocab_dict
+
+
+def make_dataset(seed=42):
+    # Pre-processing dataset
+    import re
+
+    def replace_strange_characters(text):
+        text = re.sub('[áàâ]', 'a', text)
+        text = re.sub('[ä]', 'æ', text)
+        text = re.sub('[éèëê]', 'e', text)
+        text = re.sub('[íìïî]', 'i', text)
+        text = re.sub('[óòöô]', 'o', text)
+        text = re.sub('[ö]', 'ø', text)
+        text = re.sub('[ç]', 'c', text)
+        text = re.sub('[úùüû]', 'u', text)
+        text = re.sub('\*', '', text)
+        return text
+
+    def replace_hesitations(text):
+        # text = re.sub("<[^>]*>", " ", text)  # <ee>, <qq>, <mm>, <inaudible> for NPSC. <eeeh>, <mmm> for NST-hesitate
+        text = re.sub("<ee(eh)?>", "ĥ", text)
+        text = re.sub("<mmm?>", "ĥ", text)
+        text = re.sub("<qq>", "ĥ", text)
+        # text = re.sub("<inaudible>", "I", text)
+        return text
+
+    def is_too_short(entry):
+        return len(entry["text"]) > len(entry["audio"]["array"]) // 320 or len(entry["text"]) <=1
+
+    def map_nst(entry):
+        text = entry["text"].lower()
+        text = text.replace("(...vær stille under dette opptaket...)", " ")
+        text = replace_hesitations(text)
+        text = replace_strange_characters(text)
+        text = re.sub('\s+', ' ', text)
+        return {"text": text.strip()}
+
+    def filter_nst(entry):
+        if is_too_short(entry):
+            return False  # Too short
+        if re.match(entry["type"], "pIW|CA"):
+            return False  # Spelling out words
+        if re.search("\d", entry["text"]):
+            return False
+        return True
+
+    def filter_npsc(entry):
+        if is_too_short(entry):
+            return False  # Too short
+        if re.search("\d", entry["text"]):
+            return False
+        if re.search("<inaudible>", entry["text"]):
+            return False
+        return True
+
+    def map_npsc(entry):
+        text = entry["transsentence_text"] if entry["sentence_language_code"].startswith("nn") else entry["text"]
+        text = text.lower()
+        text = replace_strange_characters(text)
+        text = replace_hesitations(text)
+        text = re.sub('\s+', ' ', text)
+        return {"text": text.strip()}
+
+    nst = datasets.load_dataset("NbAiLab/NST", "no-close")
+    npsc = datasets.load_dataset("NbAiLab/NPSC", "16K_mp3")
+    nsth = datasets.load_dataset("NbAiLab/NST_hesitate", "no")
+
+    nst = nst.map(map_nst).filter(filter_nst)
+    npsc = npsc.map(map_npsc).filter(filter_npsc)
+    nsth = nsth.map(map_nst).filter(filter_npsc)
+
+    split = len(npsc["train"]) / (len(npsc["train"]) + len(npsc["validation"]))  # Use same train/val ratio as NPSC
+    nst_train = nst["train"].train_test_split(train_size=split, seed=seed)
+    nst["train"] = nst_train["train"]
+    nst["validation"] = nst_train["test"]
+
+    nsth_train = nsth["train"].train_test_split(train_size=split, seed=seed)
+    nsth["train"] = nsth_train["train"]
+    nsth["validation"] = nsth_train["test"]
+
+    nst_base = nst.remove_columns([col for col in nst["train"].column_names if col not in ["text", "audio"]])
+    npsc_base = npsc.remove_columns([col for col in npsc["train"].column_names if col not in ["text", "audio"]])
+    nsth_base = nsth.remove_columns([col for col in nsth["train"].column_names if col not in ["text", "audio"]])
+
+    combined = {}
+    for split in "train", "validation", "test":
+        # Weight by number of examples
+        probs = np.array([len(nst_base[split]), len(npsc_base[split]), len(nsth_base[split])])
+        probs = (probs / probs.sum()).tolist()
+        comb = datasets.interleave_datasets([nst_base[split], npsc_base[split], nsth_base[split]],
+                                            probabilities=probs, seed=seed)
+        combined[split] = comb
+
+    return datasets.DatasetDict(**combined)
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Detecting last checkpoint.
+    last_checkpoint = None
+    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
+        last_checkpoint = get_last_checkpoint(training_args.output_dir)
+        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
+            raise ValueError(
+                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
+                "Use --overwrite_output_dir to overcome."
+            )
+        elif last_checkpoint is not None:
+            logger.info(
+                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
+                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
+            )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 1. First, let's load the dataset
+    raw_datasets = make_dataset(seed=training_args.seed)
+
+    if training_args.do_train:
+        if data_args.audio_column_name not in raw_datasets["train"].column_names:
+            raise ValueError(
+                f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
+                "Make sure to set `--audio_column_name` to the correct audio column - one of "
+                f"{', '.join(raw_datasets['train'].column_names)}."
+            )
+
+        if data_args.text_column_name not in raw_datasets["train"].column_names:
+            raise ValueError(
+                f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
+                "Make sure to set `--text_column_name` to the correct text column - one of "
+                f"{', '.join(raw_datasets['train'].column_names)}."
+            )
+
+        if data_args.max_train_samples is not None:
+            raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
+
+    if training_args.do_eval:
+        if data_args.max_eval_samples is not None:
+            raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
+
+    # 2. We remove some special characters from the datasets
+    # that make training complicated and do not help in transcribing the speech
+    # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
+    # that could be easily picked up by the model
+    # chars_to_ignore_regex = (
+    #    f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
+    # )
+    chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\–\_\\\+\#\/]'
+
+    text_column_name = data_args.text_column_name
+
+    def remove_special_characters(batch):
+        if chars_to_ignore_regex is not None:
+            batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
+        else:
+            batch["target_text"] = batch[text_column_name].lower() + " "
+        return batch
+
+    with training_args.main_process_first(desc="dataset map special characters removal"):
+        raw_datasets = raw_datasets.map(
+            remove_special_characters,
+            remove_columns=[text_column_name],
+            desc="remove special characters from datasets",
+        )
+
+    # save special tokens for tokenizer
+    word_delimiter_token = data_args.word_delimiter_token
+    unk_token = data_args.unk_token
+    pad_token = data_args.pad_token
+
+    # 3. Next, let's load the config as we might need it to create
+    # the tokenizer
+    # load config
+    config = AutoConfig.from_pretrained(
+        model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
+    )
+
+    # 4. Next, if no tokenizer file is defined,
+    # we create the vocabulary of the model by extracting all unique characters from
+    # the training and evaluation datasets
+    # We need to make sure that only first rank saves vocabulary
+    # make sure all processes wait until vocab is created
+    tokenizer_name_or_path = model_args.tokenizer_name_or_path
+    tokenizer_kwargs = {}
+    if tokenizer_name_or_path is None:
+        # save vocab in training output dir
+        tokenizer_name_or_path = training_args.output_dir
+
+        vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
+
+        with training_args.main_process_first():
+            if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
+                os.remove(vocab_file)
+
+        with training_args.main_process_first(desc="dataset map vocabulary creation"):
+            if not os.path.isfile(vocab_file):
+                os.makedirs(tokenizer_name_or_path, exist_ok=True)
+                vocab_dict = create_vocabulary_from_data(
+                    raw_datasets,
+                    word_delimiter_token=word_delimiter_token,
+                    unk_token=unk_token,
+                    pad_token=pad_token,
+                )
+
+                # save vocab dict to be loaded into tokenizer
+                with open(vocab_file, "w") as file:
+                    json.dump(vocab_dict, file)
+
+        # if tokenizer has just been created
+        # it is defined by `tokenizer_class` if present in config else by `model_type`
+        tokenizer_kwargs = {
+            "config": config if config.tokenizer_class is not None else None,
+            "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
+            "unk_token": unk_token,
+            "pad_token": pad_token,
+            "word_delimiter_token": word_delimiter_token,
+        }
+
+    # 5. Now we can instantiate the feature extractor, tokenizer and model
+    # Note for distributed training, the .from_pretrained methods guarantee that only
+    # one local process can concurrently download model & vocab.
+
+    # load feature_extractor and tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(
+        tokenizer_name_or_path,
+        use_auth_token=data_args.use_auth_token,
+        **tokenizer_kwargs,
+    )
+    feature_extractor = AutoFeatureExtractor.from_pretrained(
+        model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
+    )
+
+    # adapt config
+    config.update(
+        {
+            "feat_proj_dropout": model_args.feat_proj_dropout,
+            "attention_dropout": model_args.attention_dropout,
+            "hidden_dropout": model_args.hidden_dropout,
+            "final_dropout": model_args.final_dropout,
+            "mask_time_prob": model_args.mask_time_prob,
+            "mask_time_length": model_args.mask_time_length,
+            "mask_feature_prob": model_args.mask_feature_prob,
+            "mask_feature_length": model_args.mask_feature_length,
+            "gradient_checkpointing": training_args.gradient_checkpointing,
+            "layerdrop": model_args.layerdrop,
+            "ctc_loss_reduction": model_args.ctc_loss_reduction,
+            "ctc_zero_infinity": model_args.ctc_zero_infinity,
+            "pad_token_id": tokenizer.pad_token_id,
+            "vocab_size": len(tokenizer),
+            "activation_dropout": model_args.activation_dropout,
+        }
+    )
+
+    # create model
+    model = AutoModelForCTC.from_pretrained(
+        model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        config=config,
+        use_auth_token=data_args.use_auth_token,
+    )
+
+    # freeze encoder
+    if model_args.freeze_feature_encoder:
+        model.freeze_feature_encoder()
+
+    # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
+    # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
+    # so that we just need to set the correct target sampling rate and normalize the input
+    # via the `feature_extractor`
+
+    # make sure that dataset decodes audio with correct sampling rate
+    dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
+    if dataset_sampling_rate != feature_extractor.sampling_rate:
+        raw_datasets = raw_datasets.cast_column(
+            data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
+        )
+
+    # derive max & min input length for sample rate & max duration
+    max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
+    min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
+    audio_column_name = data_args.audio_column_name
+    num_workers = data_args.preprocessing_num_workers
+
+    # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
+    phoneme_language = data_args.phoneme_language
+
+    # Preprocessing the datasets.
+    # We need to read the audio files as arrays and tokenize the targets.
+    def prepare_dataset(batch):
+        # load audio
+        sample = batch[audio_column_name]
+
+        inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
+        batch["input_values"] = inputs.input_values[0]
+        batch["input_length"] = len(batch["input_values"])
+
+        # encode targets
+        additional_kwargs = {}
+        if phoneme_language is not None:
+            additional_kwargs["phonemizer_lang"] = phoneme_language
+
+        batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
+        return batch
+
+    with training_args.main_process_first(desc="dataset map preprocessing"):
+        vectorized_datasets = raw_datasets.map(
+            prepare_dataset,
+            remove_columns=next(iter(raw_datasets.values())).column_names,
+            num_proc=num_workers,
+            desc="preprocess datasets",
+        )
+
+        def is_audio_in_length_range(length):
+            return length > min_input_length and length < max_input_length
+
+        # filter data that is shorter than min_input_length
+        vectorized_datasets = vectorized_datasets.filter(
+            is_audio_in_length_range,
+            num_proc=num_workers,
+            input_columns=["input_length"],
+        )
+
+    # 7. Next, we can prepare the training.
+    # Let's use word error rate (WER) as our evaluation metric,
+    # instantiate a data collator and the trainer
+
+    # Define evaluation metrics during training, *i.e.* word error rate, character error rate
+    eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
+
+    # for large datasets it is advised to run the preprocessing on a
+    # single machine first with ``args.preprocessing_only`` since there will mostly likely
+    # be a timeout when running the script in distributed mode.
+    # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
+    # cached dataset
+    if data_args.preprocessing_only:
+        logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
+        return
+
+    def compute_metrics(pred):
+        pred_logits = pred.predictions
+        pred_ids = np.argmax(pred_logits, axis=-1)
+
+        pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
+
+        pred_str = tokenizer.batch_decode(pred_ids)
+        # we do not want to group tokens when computing the metrics
+        label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
+
+        metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
+
+        return metrics
+
+    # Now save everything to be able to create a single processor later
+    if is_main_process(training_args.local_rank):
+        # save feature extractor, tokenizer and config
+        feature_extractor.save_pretrained(training_args.output_dir)
+        tokenizer.save_pretrained(training_args.output_dir)
+        config.save_pretrained(training_args.output_dir)
+
+    try:
+        processor = AutoProcessor.from_pretrained(training_args.output_dir)
+    except (OSError, KeyError):
+        warnings.warn(
+            "Loading a processor from a feature extractor config that does not"
+            " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
+            " attribute to your `preprocessor_config.json` file to suppress this warning: "
+            " `'processor_class': 'Wav2Vec2Processor'`",
+            FutureWarning,
+        )
+        processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
+
+    # Instantiate custom data collator
+    data_collator = DataCollatorCTCWithPadding(processor=processor)
+
+    # Initialize Trainer
+    trainer = Trainer(
+        model=model,
+        data_collator=data_collator,
+        args=training_args,
+        compute_metrics=compute_metrics,
+        train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
+        eval_dataset=vectorized_datasets["validation"] if training_args.do_eval else None,
+        tokenizer=feature_extractor,
+    )
+
+    # 8. Finally, we can start training
+
+    # Training
+    if training_args.do_train:
+
+        # use last checkpoint if exist
+        if last_checkpoint is not None:
+            checkpoint = last_checkpoint
+        elif os.path.isdir(model_args.model_name_or_path):
+            checkpoint = model_args.model_name_or_path
+        else:
+            checkpoint = None
+
+        train_result = trainer.train(resume_from_checkpoint=checkpoint)
+        trainer.save_model()
+
+        metrics = train_result.metrics
+        max_train_samples = (
+            data_args.max_train_samples
+            if data_args.max_train_samples is not None
+            else len(vectorized_datasets["train"])
+        )
+        metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
+
+        trainer.log_metrics("train", metrics)
+        trainer.save_metrics("train", metrics)
+        trainer.save_state()
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+        metrics = trainer.evaluate()
+        max_eval_samples = (
+            data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
+        )
+        metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
+
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # Write model card and (optionally) push to hub
+    config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
+    kwargs = {
+        "finetuned_from": model_args.model_name_or_path,
+        "tasks": "speech-recognition",
+        "tags": ["automatic-speech-recognition", data_args.dataset_name],
+        "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
+        "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
+    }
+    if "common_voice" in data_args.dataset_name:
+        kwargs["language"] = config_name
+
+    if training_args.push_to_hub:
+        trainer.push_to_hub(**kwargs)
+    else:
+        trainer.create_model_card(**kwargs)
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "[UNK]", "pad_token": "[PAD]", "additional_special_tokens": [{"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}]}

tokenizer_config.json

@@ -0,0 +1 @@
+{"unk_token": "[UNK]", "bos_token": "<s>", "eos_token": "</s>", "pad_token": "[PAD]", "do_lower_case": false, "word_delimiter_token": "|", "replace_word_delimiter_char": " ", "special_tokens_map_file": null, "name_or_path": "./", "tokenizer_class": "Wav2Vec2CTCTokenizer", "processor_class": "Wav2Vec2ProcessorWithLM"}

train_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 15.0,
+    "train_loss": 0.17642972585578237,
+    "train_runtime": 792752.3013,
+    "train_samples": 302347,
+    "train_samples_per_second": 5.721,
+    "train_steps_per_second": 0.179
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2af229c7525c840d522bb39bbd590e1f1317ab166105438a5094178bf818b6e8
+size 3055

vocab.json

@@ -0,0 +1 @@
+{"a": 1, "b": 2, "c": 3, "d": 4, "e": 5, "f": 6, "g": 7, "h": 8, "i": 9, "j": 10, "k": 11, "l": 12, "m": 13, "n": 14, "o": 15, "p": 16, "q": 17, "r": 18, "s": 19, "t": 20, "u": 21, "v": 22, "w": 23, "x": 24, "y": 25, "z": 26, "å": 27, "æ": 28, "ø": 29, "ĥ": 30, "|": 0, "[UNK]": 31, "[PAD]": 32}