metadata

language: or
metrics:
  - wer
  - cer
tags:
  - audio
  - automatic-speech-recognition
  - speech
  - wav2vec2
  - asr
license: apache-2.0

IndicWav2Vec-Hindi

This is a Wav2Vec2 style ASR model trained in fairseq and ported to Hugging Face. More details on datasets, training-setup and conversion to HuggingFace format can be found in the IndicWav2Vec repo.

Script to Run Inference

import torch
from datasets import load_dataset
from transformers import AutoModelForCTC, AutoProcessor
import torchaudio.functional as F

DEVICE_ID = "cuda" if torch.cuda.is_available() else "cpu"
MODEL_ID = "ai4bharat/indicwav2vec-odia"

sample = next(iter(load_dataset("common_voice", "or", split="test", streaming=True)))
resampled_audio = F.resample(torch.tensor(sample["audio"]["array"]), 48000, 16000).numpy()

model = AutoModelForCTC.from_pretrained(MODEL_ID).to(DEVICE_ID)
processor = AutoProcessor.from_pretrained(MODEL_ID)

input_values = processor(resampled_audio, return_tensors="pt").input_values

with torch.no_grad():
    logits = model(input_values.to(DEVICE_ID)).logits.cpu()
    
prediction_ids = torch.argmax(logits, dim=-1)
output_str = processor.batch_decode(prediction_ids)[0]
print(f"Greedy Decoding: {output_str}")

About AI4Bharat

Website: https://ai4bharat.org/
Code: https://github.com/AI4Bharat
HuggingFace: https://huggingface.co/ai4bharat