IndicConformer 600M β€” sherpa-onnx int8, one model per language

AI4Bharat IndicConformer 600M multilingual packaged for sherpa-onnx offline recognition on CPU / Android.

The source model has one shared encoder and a 5,633-token CTC head covering 22 languages (256 tokens per language plus a blank). Run unmasked, it mixes scripts, so each file here fixes one language:

  • encoder + CTC decoder merged into a single ONNX graph,
  • a constant mask adds -inf to every token outside the chosen language's 256 (blank kept),
  • LogSoftmax output named logprobs, with the NeMo CTC metadata sherpa-onnx expects (vocab_size=5633, subsampling_factor=8, normalize_type=per_feature).

Weights are int8 (from abid1010/indic-conformer-600m-onnx). Each file is ~659 MB; all languages share indicconformer-tokens.txt.

Files

File Language
indicconformer-as.int8.onnx Assamese
indicconformer-bn.int8.onnx Bengali
indicconformer-brx.int8.onnx Bodo
indicconformer-doi.int8.onnx Dogri
indicconformer-gu.int8.onnx Gujarati
indicconformer-hi.int8.onnx Hindi
indicconformer-kn.int8.onnx Kannada
indicconformer-kok.int8.onnx Konkani
indicconformer-ks.int8.onnx Kashmiri
indicconformer-mai.int8.onnx Maithili
indicconformer-ml.int8.onnx Malayalam
indicconformer-mni.int8.onnx Manipuri
indicconformer-mr.int8.onnx Marathi
indicconformer-ne.int8.onnx Nepali
indicconformer-or.int8.onnx Odia
indicconformer-pa.int8.onnx Punjabi
indicconformer-sa.int8.onnx Sanskrit
indicconformer-sat.int8.onnx Santali
indicconformer-sd.int8.onnx Sindhi
indicconformer-ta.int8.onnx Tamil
indicconformer-te.int8.onnx Telugu
indicconformer-ur.int8.onnx Urdu
indicconformer-tokens.txt shared token table

Usage (Python)

import sherpa_onnx

rec = sherpa_onnx.OfflineRecognizer.from_nemo_ctc(
    model="indicconformer-hi.int8.onnx",
    tokens="indicconformer-tokens.txt",
    num_threads=4,
    decoding_method="greedy_search",
)
stream = rec.create_stream()
stream.accept_waveform(16000, samples)  # float32 mono, 16 kHz
rec.decode_stream(stream)
print(stream.result.text)

Input is 16 kHz mono audio. On a Galaxy S25 (4 threads) the Hindi model loads in ~1.4 s and decodes at ~0.1x real time.

Benchmark

Word error rate on synthetic medical consultation dialogues (Svara TTS, ~470 clips per language). Each recording is split into turns with Silero VAD and each turn is decoded once. Punctuation and tags are ignored; spelling variants (e.g. nukta) count as errors.

Language WER Language WER
Marathi 11.5% Malayalam 21.8%
Telugu 11.5% Nepali 24.8%
Kannada 15.4% Punjabi 26.4%
Hindi 16.0% Gujarati 27.2%
Bengali 16.6% Maithili 29.6%
Tamil 18.1% Dogri 42.9%
Sanskrit 19.7% Bodo 43.6%
Assamese 20.1%

Odia, Urdu, Konkani, Kashmiri, Manipuri, Santali and Sindhi were not benchmarked.

License and attribution

MIT, same as the source models.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Docglasses-AI/indic-conformer-600m-sherpa-onnx-int8

Quantized
(17)
this model