gigaam-multilingual-ctc-onnx-int8

GigaAM Multilingual CTC — ONNX int8 for sherpa-onnx

Self-exported ONNX build of the GigaAM Multilingual ASR model (revision ctc, 220M Conformer, charwise CTC), dynamically quantized to int8 for CPU inference with sherpa-onnx.

The purpose of this build: high-quality Russian dictation with English technical terms transcribed in Latin script (e.g. dictating the Russian phrase for "run npm install" produces npm install, not transliterated Cyrillic), running offline on small CPU-only servers (tested on 2 vCPU / 1.9 GB RAM).

Files

Provenance

• Base model: ai-sage/GigaAM-Multilingual, revision ctc (220M)
• Export: gigaam package (model.to_onnx()), fp32 → dynamic int8 quantization
• Vocabulary note: multilingual keeps it under cfg["decoding"]["vocabulary"] (v1/v2 CTC used cfg["labels"]) — handled in the export script
• Model metadata embedded: model_type=EncDecCTCModel, subsampling_factor=4, is_giga_am=1

Build pipeline and integration into OpenChamber (web UI for the OpenCode agent) live in the companion git repository: openchamber-speech-ru (build, sanity tests, OpenChamber patch, install scripts).

Benchmarks

Word Error Rate (%) from the base model card (greedy decoding, normalized):

Measured on a 2-vCPU VPS (this int8 build, 2 threads, sherpa-onnx 1.12.28):

Usage

sherpa-onnx (Python)

import sherpa_onnx

recognizer = sherpa_onnx.OfflineRecognizer.from_nemo_ctc(
    model="model.int8.onnx",
    tokens="tokens.txt",
    num_threads=2,
    provider="cpu",
)
stream = recognizer.create_stream()
stream.accept_waveform(16000, samples)  # float32 in [-1, 1]
recognizer.decode_streams([stream])
print(stream.result.text)

sherpa-onnx (Node.js)

const sherpa = require('sherpa-onnx-node');
const recognizer = new sherpa.OfflineRecognizer({
  featConfig: { sampleRate: 16000, featureDim: 80 },
  modelConfig: {
    nemoCtc: { model: 'model.int8.onnx' },
    tokens: 'tokens.txt',
    numThreads: 2,
    provider: 'cpu',
    debug: 0,
  },
  decodingMethod: 'greedy_search',
});

OpenChamber

See the companion repository (openchamber-speech-ru): install.sh downloads this model, patches the OpenChamber dictation worker with a nemo_ctc engine branch, and activates it for the built-in microphone dictation.

Limitations

• No punctuation, no capitalization (charwise CTC head)
• English speech recognition is moderate (FLEURS WER ~12); embedded English terms inside Russian speech come out in Latin script and are usually accurate, but long English dictation is better served by a Whisper-family model
• Russian-only quality is slightly below GigaAM v2/v3 RNNT (FLEURS 4.4 vs ~3), which is the trade-off for multilingual output

License

MIT — inherited from the base model, Copyright (c) 2024 GigaChat Team (salute-developers/GigaAM). The export script is adapted from k2-fsa/sherpa-onnx (Apache-2.0).

Citation

@misc{gigaam_multilingual,
  title={GigaAM Multilingual: Foundation Model for Underrepresented Languages},
  author={Kuzmenko, Andrei and Maximenko, Alexandr and Kutsakov, Aleksandr and Gospodinov, Georgii and Bolotov, Dmitrii and Bogomolov, Pavel and Minkin, Fyodor},
  year={2026},
  eprint={2607.10371},
  archivePrefix={arXiv},
}