README.md

metadata

language: mr
datasets:
  - openslr
metrics:
  - wer
tags:
  - audio
  - automatic-speech-recognition
  - speech
  - xlsr-fine-tuning-week
license: apache-2.0
model-index:
  - name: XLSR Wav2Vec2 Large 53 Marathi by Sumedh Khodke
    results:
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: OpenSLR mr
          type: openslr
        metrics:
          - name: Test WER
            type: wer
            value: 12.7

Wav2Vec2-Large-XLSR-53-Marathi

Fine-tuned facebook/wav2vec2-large-xlsr-53 on Marathi using the OpenSLR SLR64 dataset. When using this model, make sure that your speech input is sampled at 16kHz. This data contains only female voices but it works well for male voices too. WER (Word Error Rate) on the Test Set: 12.70 %

Usage

The model can be used directly without a language model as follows, given that your dataset has Marathi actual_text and path_in_folder columns:

import torch, torchaudio
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

mr_test_dataset_new = all_data['test']

processor = Wav2Vec2Processor.from_pretrained("sumedh/wav2vec2-large-xlsr-marathi") 
model = Wav2Vec2ForCTC.from_pretrained("sumedh/wav2vec2-large-xlsr-marathi") 

resampler = torchaudio.transforms.Resample(48_000, 16_000) #first arg - input sample, second arg - output sample
# Preprocessing the datasets. We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
  speech_array, sampling_rate = torchaudio.load(batch["path_in_folder"])
  batch["speech"] = resampler(speech_array).squeeze().numpy()
  return batch
mr_test_dataset_new = mr_test_dataset_new.map(speech_file_to_array_fn)
inputs = processor(mr_test_dataset_new["speech"][:5], sampling_rate=16_000, return_tensors="pt", padding=True)
with torch.no_grad():
  logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
predicted_ids = torch.argmax(logits, dim=-1)
print("Prediction:", processor.batch_decode(predicted_ids))
print("Reference:", mr_test_dataset_new["actual_text"][:5])

Evaluation

Evaluated on 10% of the Marathi data on Open SLR-64.

import re, torch, torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

mr_test_dataset_new = all_data['test']
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("sumedh/wav2vec2-large-xlsr-marathi")
model = Wav2Vec2ForCTC.from_pretrained("sumedh/wav2vec2-large-xlsr-marathi") 
model.to("cuda")

chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“]' 
resampler = torchaudio.transforms.Resample(48_000, 16_000)
# Preprocessing the datasets. We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
  batch["actual_text"] = re.sub(chars_to_ignore_regex, '', batch["actual_text"]).lower()
  speech_array, sampling_rate = torchaudio.load(batch["path_in_folder"])
  batch["speech"] = resampler(speech_array).squeeze().numpy()
  return batch
mr_test_dataset_new = mr_test_dataset_new.map(speech_file_to_array_fn)
def evaluate(batch):
  inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
  with torch.no_grad():
    logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
    pred_ids = torch.argmax(logits, dim=-1)
    batch["pred_strings"] = processor.batch_decode(pred_ids)
  return batch
result = mr_test_dataset_new.map(evaluate, batched=True, batch_size=8)
print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["actual_text"])))

Training

Train-Test ratio was 90:10. Colab training notebook can be found here.

Training Config and Summary

weights-and-biases run summary here