Fine-tuned facebook/wav2vec2-large-xlsr-53 Bengali using a subset of 40,000 utterances from Bengali ASR training data set containing ~196K utterances. Tested WER using ~4200 held out from training. When using this model, make sure that your speech input is sampled at 16kHz. Train Script can be Found at : train.py

Data Prep Notebook : https://colab.research.google.com/drive/1JMlZPU-DrezXjZ2t7sOVqn7CJjZhdK2q?usp=sharing
Inference Notebook : https://colab.research.google.com/drive/1uKC2cK9JfUPDTUHbrNdOYqKtNozhxqgZ?usp=sharing

Usage

The model can be used directly (without a language model) as follows:

python

import torch
import torchaudio
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

processor = Wav2Vec2Processor.from_pretrained("arijitx/wav2vec2-large-xlsr-bengali")
model = Wav2Vec2ForCTC.from_pretrained("arijitx/wav2vec2-large-xlsr-bengali")
# model = model.to("cuda")

resampler = torchaudio.transforms.Resample(TEST_AUDIO_SR, 16_000)
def speech_file_to_array_fn(batch):
    speech_array, sampling_rate = torchaudio.load(batch)
    speech =  resampler(speech_array).squeeze().numpy()
    return speech

speech_array = speech_file_to_array_fn("test_file.wav")
inputs = processor(speech_array, sampling_rate=16_000, return_tensors="pt", padding=True)
with torch.no_grad():
    logits = model(inputs.input_values).logits

    
predicted_ids = torch.argmax(logits, dim=-1)
preds = processor.batch_decode(predicted_ids)[0]
print(preds.replace("[PAD]",""))

Test Result: WER on ~4200 utterance : 32.45 %