GigaAM v3 Russian E2E CTC INT8 ONNX
Русская офлайн-модель распознавания речи GigaAM v3 E2E CTC в формате ONNX INT8, подготовленная для Sherpa-ONNX.
Russian offline ASR model based on GigaAM v3 E2E CTC, exported to ONNX and quantized to INT8. This package is ready for Sherpa-ONNX's offline NeMo CTC backend.
Highlights
- Russian ASR
- Non-autoregressive CTC decoding
- E2E output with punctuation and text normalization
- 16 kHz mono PCM input
- INT8 ONNX model, approximately 215 MiB
- Sherpa-ONNX compatible model metadata
- MIT license
Files
| File | Purpose | Size | SHA-256 |
|---|---|---|---|
model.int8.onnx |
Sherpa-ONNX compatible E2E CTC model | 224,893,668 bytes | 784d77aaa7a0899a247e72c71d667792087ee4dec985db1f57961767a5f6014f |
tokens.txt |
Token vocabulary | 2,007 bytes | 142de7570b3de5b3035ce111a89c228e80e6085273731d944093ddf24fa539cd |
Model Contract
| Property | Value |
|---|---|
| Sample rate | 16,000 Hz |
| Channels | Mono |
| Feature type | Log Mel |
| Feature dimensions | 64 |
| Mel scale | HTK |
| Window length | 320 samples / 20 ms |
| Hop length | 160 samples / 10 ms |
| FFT size | 320 |
| CTC classes | 257 |
The ONNX graph expects precomputed 64-dimensional HTK Log Mel features. It is not a zipformer2_ctc model. With Sherpa-ONNX, create an offline recognizer using the NeMo EncDec CTC backend and a 64-dimensional feature configuration.
Sherpa-ONNX Compatibility
model.int8.onnx contains the original INT8 graph and weights plus ONNX metadata required by Sherpa-ONNX:
model_type=EncDecCTCModelvocab_size=257subsampling_factor=4is_giga_am=1
No graph nodes, weights, quantization parameters, vocabulary entries, or decoding behavior were changed. The added metadata is why the packaged model is 321 bytes larger than the upstream ONNX file.
Provenance
This repository packages the following upstream ONNX export:
- Upstream repository:
istupakov/gigaam-v3-onnx - Upstream commit:
322c3b29492673eb7d0b434bfa9dfb8653e34d02 - Base model:
ai-sage/GigaAM-v3 - Export tooling:
onnx-asr
The upstream source artifacts were:
| Upstream file | SHA-256 |
|---|---|
v3_e2e_ctc.int8.onnx |
2e3fcb7a7b66030336fd10c2fcfb033bd1dc7e1bf238fe5cfd83b1d0cfc9d28e |
v3_e2e_ctc_vocab.txt |
142de7570b3de5b3035ce111a89c228e80e6085273731d944093ddf24fa539cd |
The published tokens.txt is the upstream vocabulary renamed for the Sherpa-ONNX package. The published model has only the compatibility metadata described above appended to the ONNX container.
License
This distribution is released under the MIT License, consistent with the upstream ONNX export. Review the licenses and terms of all upstream dependencies before commercial distribution.
Verification
After downloading, verify both published files:
shasum -a 256 model.int8.onnx tokens.txt
Expected results:
784d77aaa7a0899a247e72c71d667792087ee4dec985db1f57961767a5f6014f model.int8.onnx
142de7570b3de5b3035ce111a89c228e80e6085273731d944093ddf24fa539cd tokens.txt