Sber GigaAM-Multilingual (char-wise CTC head, 70+ languages) exported to the sherpa-onnx NeMo-CTC offline format. Ready to use with any sherpa-onnx binding (C/C++, Python, Go, Java, C#, ...).

Both ASR variants are provided:

220M (ctc revision, 16-layer Conformer, d_model 768) — repo root
600M (large_ctc revision, 24-layer Conformer, d_model 1024) — large/ subfolder; noticeably more accurate (especially on English), ~2.5× slower on CPU (int8: RTF ≈ 0.14 vs ≈ 0.06 for 220M, Apple silicon, 4 threads)

The export follows the same recipe as the official sherpa-onnx GigaAM v2/v3 CTC packs (export-onnx-ctc-multilingual.py in this repo is an adaptation of that pack's export script): sherpa metadata (model_type=EncDecCTCModel, is_giga_am=1, subsampling_factor=4, vocab_size=71) is embedded into the ONNX graph and the char-wise vocabulary is written as tokens.txt (space, ', a–z, Cyrillic incl. ё, and Kazakh/Kyrgyz letters + <blk>).

Files

File	Description
`model.int8.onnx`	220M, dynamic-quantized int8 (225 MB) — recommended for CPU
`model.onnx`	220M, fp32 export (885 MB)
`tokens.txt`	71 char-wise tokens (`<token> <id>`, `<blk>` last) — identical for both variants
`multilingual_ctc.yaml`	source model config from the `gigaam` package
`large/model.int8.onnx`	600M, int8, single self-contained file (592 MB)
`large/model.onnx` + `large/model.onnx.data`	600M fp32; weights exceed the 2 GB protobuf limit, so the graph references one external-data file (must be kept side by side)
`large/tokens.txt`, `large/multilingual_large_ctc.yaml`	tokens (same 71) and source config of the large variant
`export-onnx-ctc-multilingual.py`	reproducible export script (`pip install gigaam onnx onnxruntime`), `--model multilingual_ctc \| multilingual_large_ctc`

Feature extraction

16 kHz mono, 64 mel bins, n_fft=320, win_length=320, hop_length=160 — i.e. feature_dim=64 in sherpa-onnx (FeatConfig); everything else is read from the embedded metadata.

Usage

Python

python

import sherpa_onnx

recognizer = sherpa_onnx.OfflineRecognizer.from_nemo_ctc(
    model="model.int8.onnx",
    tokens="tokens.txt",
    num_threads=4,
    sample_rate=16000,
    feature_dim=64,
)
stream = recognizer.create_stream()
stream.accept_waveform(16000, samples)  # float32 mono 16 kHz
recognizer.decode_stream(stream)
print(stream.result.text)

Go

config := &sherpa.OfflineRecognizerConfig{}
config.FeatConfig.SampleRate = 16000
config.FeatConfig.FeatureDim = 64
config.ModelConfig.NemoCTC.Model = "model.int8.onnx"
config.ModelConfig.Tokens = "tokens.txt"
config.ModelConfig.ModelType = "nemo_ctc"
config.DecodingMethod = "greedy_search"
recognizer := sherpa.NewOfflineRecognizer(config)

Notes and limitations

Output is lowercase without punctuation (char-wise CTC head).
Hotwords / contextual biasing are not available: sherpa-onnx supports hotwords only for transducer models, and GigaAM-Multilingual ships CTC only.
Informal benchmark on an Apple-silicon CPU (4 threads, int8): RTF ≈ 0.06. On noisy Russian meeting audio quality is on par with GigaAM v3; English is usable but noticeably weaker than English-centric models (matches the upstream Common Voice/FLEURS numbers).
Word-level timestamps are produced (per-token stamps, words split on the space token).

License

MIT, same as the upstream GigaAM repository (see LICENSE). All model weights are © SberDevices / salute-developers — this repo only repackages them for the sherpa-onnx runtime.

fussraider/GigaAM-Multilingual-sherpa-onnx-ctc

GigaAM-Multilingual CTC for sherpa-onnx