README.md

metadata

language: hi
metrics:
  - wer
tags:
  - audio
  - automatic-speech-recognition
  - speech
  - xlsr-fine-tuning-week
license: cc
model-index:
  - name: Wav2Vec2 Hindi Model by Swayam Mittal
    results:
      - task:
          name: Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice hi
          type: common_voice
          args: hi
        metrics:
          - name: Test WER
            type: wer
            value: 24.17

hindi-clsril-100

Fine-tuned Harveenchadha/wav2vec2-pretrained-clsril-23-10k on Hindi using the Common Voice, included openSLR Hindi dataset. When using this model, make sure that your speech input is sampled at 16kHz.

Evaluation

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

#!pip install datasets==1.4.1
#!pip install transformers==4.4.0
#!pip install torchaudio
#!pip install jiwer

import torch
import torchaudio
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2ProcessorWithLM
processor_with_lm = Wav2Vec2ProcessorWithLM.from_pretrained("swayam01/hindi-clsril-100")
model = Wav2Vec2ForCTC.from_pretrained("swayam01/hindi-clsril-100")
test_dataset = load_dataset("common_voice", "hi", split="test") 
wer = load_metric("wer")
chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\�\।\']'
resampler = torchaudio.transforms.Resample(48_000, 16_000)
# We need to read the audio 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)

def evaluate(batch):
    inputs = processor_with_lm(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
    batch["pred_strings"] = transcription = processor_with_lm.batch_decode(logits.numpy()).text
    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: 24.17 %

Training

The Common Voice hi train, validation were used for training, as well as openSLR hi train, validation and test datasets.

The script used for training can be found here colab