Wav2vec2 Large Xlsr Bengali

🚀 Wav2Vec2-Large-XLSR-Bengali

This model is a fine - tuned version of facebook/wav2vec2-large-xlsr-53 for Bengali. It uses a subset of 40,000 utterances from the Bengali ASR training data set containing ~196K utterances. The Word Error Rate (WER) is tested using ~4200 utterances held out from training.

🚀 Quick Start

When using this model, make sure that your speech input is sampled at 16kHz. The train script can be found at train.py.

• Data Preparation Notebook: Link
• Inference Notebook: Link

✨ Features

• Language: Bengali
• Datasets: OpenSLR
• Metrics: Word Error Rate (WER)
• Tags: bn, audio, automatic-speech-recognition, speech
• License: CC BY - SA 4.0

📦 Installation

No specific installation steps are provided in the original document.

💻 Usage Examples

Basic Usage

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

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 %

📚 Documentation

The model is fine - tuned on a subset of Bengali ASR training data. The test WER is calculated using held - out data. When using the model, ensure the input speech is sampled at 16kHz.

🔧 Technical Details

The model is a fine - tuned version of facebook/wav2vec2-large-xlsr-53 for Bengali. It uses a subset of 40,000 utterances from the Bengali ASR training data set. The test WER is 32.45% on ~4200 held - out utterances.

📄 License

This model is released under the CC BY - SA 4.0 license.

Model Index