Vaani (ARTPARK-IISc/Vaani-transcription-part): 121,960 training samples across 7 languages
Proprietary MWire corpus: 28,524 training samples across 5 languages (Khasi, Garo, Mizo, Nagamese, Kokborok)
Total: 150,483 training samples

Training Details

Base model: openai/whisper-medium
Learning rate: 1e-5 with 500 warmup steps
Steps: 8,000
Batch size: 16 (gradient accumulation 2, effective 32)
Mixed precision: fp16
Language token: welsh proxy for all languages except Assamese

Usage

python

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import soundfile as sf
import torch

processor = WhisperProcessor.from_pretrained("MWirelabs/ne-asr")
model = WhisperForConditionalGeneration.from_pretrained("MWirelabs/ne-asr")

audio, sr = sf.read("audio.wav")
inputs = processor(audio, sampling_rate=16000, return_tensors="pt")

# Force language token (use "welsh" for all except Assamese)
forced_ids = processor.get_decoder_prompt_ids(language="welsh", task="transcribe")
with torch.no_grad():
    predicted_ids = model.generate(inputs.input_features, forced_decoder_ids=forced_ids)

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

Citation

If you use this model, please cite:

@inproceedings{nyalang2026nemultispeech,
title={{NE}-MultiSpeech: A Multilingual Speech Corpus and {ASR} Benchmark for Northeast Indian Languages},
author={Nyalang, Badal and Borah, Angana},
booktitle={The 2026 Conference on Empirical Methods in Natural Language Processing},
year={2026},
url={https://openreview.net/forum?id=cOCsA88Fza}
}

License

CC-BY-4.0. Developed by MWire Labs, Shillong, Meghalaya.