IndicConformer
AI4Bharat's IndicConformers is a suite of ASR models built to deliver accurate speech-to-text conversion in all 22 official Indian languages. By leveraging cutting-edge deep learning techniques, these models provide precise transcriptions. As the country's first open-source ASR system covering such a vast array of languages, AI4Bharat Indic Conformer is a transformative tool for making technology more inclusive and accessible to all. IndicConformer is released under the MIT license.
Model Details
- Model Name: IndicConformer-600M-Multi
- Repository: ai4bharat/indic-conformer-600m-multilingual
- Architecture: Multilingual Conformer-based Hybrid CTC + RNNT ASR model
- Parameter Size: 600M
- Languages Supported: IN-22
🚩NOTE 🚩Indic-Transcribe is now available. It supports 25 Indian languages and has better transcription quality compared to IndicConformer. We recommend using Indic-Transcribe.
Model Usage
This model can be used to transcribe speech in various Indian languages. It supports two decoding strategies:
- CTC (Connectionist Temporal Classification)
- RNNT (Recurrent Neural Network Transducer)
Installation
Ensure that you have transformers and torchaudio installed:
pip install transformers torchaudio "onnxruntime==1.20.1" "onnx==1.20.1" "onnxruntime-gpu==1.20.1"
Inference Example
from transformers import AutoModel
import torch, torchaudio
# Load the model
model = AutoModel.from_pretrained("ai4bharat/indic-conformer-600m-multilingual", trust_remote_code=True)
# Load an audio file
wav, sr = torchaudio.load("audio.flac")
wav = torch.mean(wav, dim=0, keepdim=True)
target_sample_rate = 16000 # Expected sample rate
if sr != target_sample_rate:
resampler = torchaudio.transforms.Resample(orig_freq=sr, new_freq=target_sample_rate)
wav = resampler(wav)
# Perform ASR with CTC decoding
transcription_ctc = model(wav, "hi", "ctc")
print("CTC Transcription:", transcription_ctc)
# Perform ASR with RNNT decoding
transcription_rnnt = model(wav, "hi", "rnnt")
print("RNNT Transcription:", transcription_rnnt)
Supported Languages
IndicConformer-600M-Multi is trained for 22 officially recognized languages of India, including:
- Assamese(
as) - Bengali(
bn) - Bodo(
brx) - Dogri(
doi) - Gujarati(
gu) - Hindi(
hi) - Kannada(
kn) - Konkani(
kok) - Kashmiri(
ks) - Maithili(
mai) - Malayalam(
ml) - Manipuri(
mni) - Marathi(
mr) - Nepali(
ne) - Odia(
or) - Punjabi(
pa) - Sanskrit(
sa) - Santali(
sat) - Sindhi(
sd) - Tamil(
ta) - Telugu(
te) - Urdu(
ur)
The tokenizers used for these languages can be found here
Contact
For any questions or feedback, please contact:
- Tahir Javed (tahir@cse.iitm.ac.in)
- Kaushal Bhogale (CS22D006@cse.iitm.ac.in)