vakyansh-wav2vec2-hindi-him-4200

Note: The result from this model is without a language model so you may witness a higher WER in some cases.

Dataset

This model was trained on 4200 hours of Hindi Labelled Data. The labelled data is not present in public domain as of now.

Training Script

Models were trained using experimental platform setup by Vakyansh team at Ekstep. Here is the training repository.

In case you want to explore training logs on wandb they are here.

Usage

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

import soundfile as sf
import torch
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import argparse

def parse_transcription(wav_file):
    # load pretrained model
    processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-hindi-him-4200")
    model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-hindi-him-4200")

    # load audio
    audio_input, sample_rate = sf.read(wav_file)

    # pad input values and return pt tensor
    input_values = processor(audio_input, sampling_rate=sample_rate, return_tensors="pt").input_values

    # INFERENCE
    # retrieve logits & take argmax
    logits = model(input_values).logits
    predicted_ids = torch.argmax(logits, dim=-1)

    # transcribe
    transcription = processor.decode(predicted_ids[0], skip_special_tokens=True)
    print(transcription)

Evaluation

The model can be evaluated as follows on the hindi test data of Common Voice.

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

test_dataset = load_dataset("common_voice", "hi", split="test")
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("Harveenchadha/vakyansh-wav2vec2-hindi-him-4200")
model = Wav2Vec2ForCTC.from_pretrained("Harveenchadha/vakyansh-wav2vec2-hindi-him-4200")
model.to("cuda")

resampler = torchaudio.transforms.Resample(48_000, 16_000)

chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“]'

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
  batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()
  speech_array, sampling_rate = torchaudio.load(batch["path"])
  batch["speech"] = resampler(speech_array).squeeze().numpy()
  return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
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")).logits

      pred_ids = torch.argmax(logits, dim=-1)
      batch["pred_strings"] = processor.batch_decode(pred_ids, skip_special_tokens=True)
      return batch

result = test_dataset.map(evaluate, batched=True, batch_size=8)

print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))

Test Result: 33.17 %

Credits

Thanks to Ekstep Foundation for making this possible. The vakyansh team will be open sourcing speech models in all the Indic Languages.