zeroshot.py

import os
import tempfile
import re
import librosa
import torch
import json
import numpy as np

from transformers import Wav2Vec2ForCTC, AutoProcessor
from huggingface_hub import hf_hub_download
from torchaudio.models.decoder import ctc_decoder
from utils.text_norm import text_normalize
from utils.lm import create_unigram_lm, maybe_generate_pseudo_bigram_arpa

uroman_dir = "uroman"
assert os.path.exists(uroman_dir)
UROMAN_PL = os.path.join(uroman_dir, "bin", "uroman.pl")

ASR_SAMPLING_RATE = 16_000

WORD_SCORE_DEFAULT_IF_LM = -0.18
WORD_SCORE_DEFAULT_IF_NOLM = -3.5
LM_SCORE_DEFAULT = 1.48

MODEL_ID = "upload/mms_zs"

processor = AutoProcessor.from_pretrained(MODEL_ID)
model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)

token_file = "upload/mms_zs/tokens.txt"


class MY_LOG:
    def __init__(self):
        self.text = "[START]"

    def add(self, new_log, new_line= True):
        self.text = self.text + ("\n" if new_line else " ") + new_log
        self.text = self.text.strip()
        return self.text


def error_check_file(filepath):
    if not isinstance(filepath, str):
        return "Expected file to be of type 'str'. Instead got {}".format(
            type(filepath)
        )
    if not os.path.exists(filepath):
        return "Input file '{}' doesn't exists".format(type(filepath))


def norm_uroman(text):
    text = text.lower()
    text = text.replace("’", "'")
    text = re.sub("([^a-z' ])", " ", text)
    text = re.sub(" +", " ", text)
    return text.strip()


def uromanize(words):
    iso = "xxx"
    with tempfile.NamedTemporaryFile() as tf, tempfile.NamedTemporaryFile() as tf2:
        with open(tf.name, "w") as f:
            f.write("\n".join(words))
        cmd = f"perl " + UROMAN_PL
        cmd += f" -l {iso} "
        cmd += f" < {tf.name} > {tf2.name}"
        os.system(cmd)
        lexicon = {}
        with open(tf2.name) as f:
            for idx, line in enumerate(f):
                if not line.strip():
                    continue
                line = re.sub(r"\s+", "", norm_uroman(line)).strip()
                lexicon[words[idx]] = " ".join(line) + " |"
    return lexicon


def filter_lexicon(lexicon, word_counts):
    spelling_to_words = {}
    for w, s in lexicon.items():
        spelling_to_words.setdefault(s, [])
        spelling_to_words[s].append(w)

    lexicon = {}
    for s, ws in spelling_to_words.items():
        if len(ws) > 1:
            # use the word which has higest counts, fewed additional characters
            ws.sort(key=lambda w: (-word_counts[w], len(w)))
        lexicon[ws[0]] = s
    return lexicon


def load_words(filepath):
    words = {}
    num_sentences = 0
    with open(filepath) as f:
        for line in f:
            line = line.strip().lower()
            num_sentences += 1
            line = text_normalize(line, iso_code="xxx")
            # ignore invalid words.
            for w in line.split():
                words.setdefault(w, 0)
                words[w] += 1
    return words, num_sentences


def process(
    audio_data,
    words_file,
    lm_path=None,
    wscore=None,
    lmscore=None,
    wscore_usedefault=True,
    lmscore_usedefault=True,
    autolm=True,
    reference=None,
):
    transcription, logs = "", MY_LOG()
    if not audio_data or not words_file:
        yield "ERROR: Empty audio data or words file", logs.text
        return
    if isinstance(audio_data, tuple):
        # microphone
        sr, audio_samples = audio_data
        audio_samples = (audio_samples / 32768.0).astype(float)

        if sr != ASR_SAMPLING_RATE:
            audio_samples = librosa.resample(audio_samples, sr, ASR_SAMPLING_RATE)
    else:
        # file upload
        assert isinstance(audio_data, str)
        audio_samples = librosa.load(audio_data, sr=ASR_SAMPLING_RATE, mono=True)[0]
    yield transcription, logs.add(f"Number of audio samples: {len(audio_samples)}")

    inputs = processor(
        audio_samples, sampling_rate=ASR_SAMPLING_RATE, return_tensors="pt"
    )

    # set device
    if torch.cuda.is_available():
        device = torch.device("cuda")
    elif (
        hasattr(torch.backends, "mps")
        and torch.backends.mps.is_available()
        and torch.backends.mps.is_built()
    ):
        device = torch.device("mps")
    else:
        device = torch.device("cpu")
    device = torch.device("cpu")
    model.to(device)
    inputs = inputs.to(device)
    yield transcription, logs.add(f"Using device: {device}")

    with torch.no_grad():
        outputs = model(**inputs).logits

    # Setup lexicon and decoder
    yield transcription, logs.add(f"Loading words....")
    try:
        word_counts, num_sentences = load_words(words_file)
    except Exception as e:
        yield f"ERROR: Loading words failed '{str(e)}'", logs.text
        return

    yield transcription, logs.add(
        f"Loaded {len(word_counts)} words from {num_sentences} lines.\nPreparing lexicon...."
    )

    try:
        lexicon = uromanize(list(word_counts.keys()))
    except Exception as e:
        yield f"ERROR: Creating lexicon failed '{str(e)}'", logs.text
        return
    # for k, v in lexicon.items():
    #     if len(v) < 5:
    #         print(k, v)
    yield transcription, logs.add(f"Leixcon size: {len(lexicon)}")

    # Input could be sentences OR list of words. Check if atleast one word has a count > 1 to diffentiate 
    tmp_file = tempfile.NamedTemporaryFile() # could be used for LM
    if autolm and any([cnt > 2 for cnt in word_counts.values()]):
        yield transcription, logs.add(f"Creating unigram LM...", False)
        lm_path = tmp_file.name 
        create_unigram_lm(word_counts, num_sentences, lm_path)
        yield transcription, logs.add(f"OK")

    
    if lm_path is None:
        yield transcription, logs.add(f"Filtering lexicon....")
        lexicon = filter_lexicon(lexicon, word_counts)
        yield transcription, logs.add(
            f"Ok. Leixcon size after filtering: {len(lexicon)}"
        )
    else:
        # kenlm throws an error if unigram LM is being used  
        # HACK: generate a bigram LM from unigram LM and a dummy bigram to trick it 
        maybe_generate_pseudo_bigram_arpa(lm_path)

    # for k, v in lexicon.items():
    #     if len(v) < 5:
    #         print(k, v)

    # print(lexicon["the"], lexicon["\"(t)he"])
    with tempfile.NamedTemporaryFile() as lexicon_file:
        if lm_path is not None and not lm_path.strip():
            lm_path = None

        with open(lexicon_file.name, "w") as f:
            idx = 10
            for word, spelling in lexicon.items():
                f.write(word + " " + spelling + "\n")
                idx += 1

        if wscore_usedefault:
            wscore = (
                WORD_SCORE_DEFAULT_IF_LM
                if lm_path is not None
                else WORD_SCORE_DEFAULT_IF_NOLM
            )
        if lmscore_usedefault:
            lmscore = LM_SCORE_DEFAULT if lm_path is not None else 0

        yield transcription, logs.add(
            f"Using word score: {wscore}\nUsing lm score: {lmscore}"
        )

        beam_search_decoder = ctc_decoder(
            lexicon=lexicon_file.name,
            tokens=token_file,
            lm=lm_path,
            nbest=1,
            beam_size=500,
            beam_size_token=50,
            lm_weight=lmscore,
            word_score=wscore,
            sil_score=0,
            blank_token="<s>",
        )

        beam_search_result = beam_search_decoder(outputs.to("cpu"))
        transcription = " ".join(beam_search_result[0][0].words).strip()

    yield transcription, logs.add(f"[DONE]")


for i in process("upload/english/english.mp3", "upload/english/c4_5k_sentences.txt"):
    print(i)


# for i in process("upload/ligurian/ligurian_1.mp3", "upload/ligurian/zenamt_5k_sentences.txt"):
#     print(i)