Anshul1212/indicconformer-hi-hybrid-rnnt-large-hf

IndicConformer Hindi ASR (HuggingFace Pipeline)

HuggingFace-compatible conversion of the AI4Bharat IndicConformer model. This enables using the model with standard HuggingFace Transformers patterns instead of NeMo.

Installation

pip install transformers torch torchaudio sentencepiece librosa

Quick Start (Pipeline API)

from huggingface_hub import snapshot_download
import sys

# Download model
model_path = snapshot_download("Anshul1212/indicconformer-hi-hybrid-rnnt-large-hf")
sys.path.insert(0, model_path)

# Import pipeline and model
from pipeline_indicconformer import IndicConformerASRPipeline
from modeling_indicconformer import IndicConformerForCTC

# Create pipeline
model = IndicConformerForCTC.from_pretrained(model_path)
pipe = IndicConformerASRPipeline(model=model, model_path=model_path)

# Transcribe
result = pipe("audio.wav")
print(result["text"])

Alternative: Auto Classes

import torch
import torchaudio
from transformers import AutoModel, AutoTokenizer, AutoFeatureExtractor

# Load components
model = AutoModel.from_pretrained("Anshul1212/indicconformer-hi-hybrid-rnnt-large-hf", trust_remote_code=True)
model.eval()

feature_extractor = AutoFeatureExtractor.from_pretrained("Anshul1212/indicconformer-hi-hybrid-rnnt-large-hf", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("Anshul1212/indicconformer-hi-hybrid-rnnt-large-hf", trust_remote_code=True)

# Load audio
waveform, sample_rate = torchaudio.load("audio.wav")

# Extract features and transcribe
inputs = feature_extractor(waveform.squeeze().numpy(), sampling_rate=16000, return_tensors="pt")

with torch.no_grad():
    predicted_ids = model.generate(
        input_features=inputs['input_features'],
        language='hi',
        decoder_mode='rnnt'
    )
    text = tokenizer.decode(predicted_ids[0])

print(text)

GPU Inference

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

input_features = inputs['input_features'].to(device)

with torch.no_grad():
    predicted_ids = model.generate(input_features=input_features, language='hi', decoder_mode='rnnt')
    text = tokenizer.decode(predicted_ids[0])

CTC Decoding

with torch.no_grad():
    predicted_ids = model.generate(
        input_features=inputs['input_features'],
        language='hi',
        decoder_mode='ctc'
    )
    text = tokenizer.decode(predicted_ids[0], use_ctc=True)

Supported Languages

Code	Language	Code	Language
hi	Hindi	te	Telugu
bn	Bengali	ta	Tamil
gu	Gujarati	ml	Malayalam
mr	Marathi	kn	Kannada
pa	Punjabi	or	Odia
as	Assamese	ur	Urdu
ne	Nepali	sa	Sanskrit
sd	Sindhi	kok	Konkani
doi	Dogri	mai	Maithili
mni	Manipuri	brx	Bodo
sat	Santali	ks	Kashmiri

Citation

@misc{indicconformer2024,
  title={IndicConformer: Conformer-based ASR for Indian Languages},
  author={AI4Bharat},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/ai4bharat/indicconformer_stt_hi_hybrid_ctc_rnnt_large}
}

License

Apache 2.0