README.md

---
license: mit
library_name: onnx
tags:
  - automatic-speech-recognition
  - speech
  - russian
  - onnx
  - rnnt
  - gigaam
language:
  - ru
base_model:
  - ai-sage/GigaAM-v3
---

# CoppersideASR

CoppersideASR is an ONNX RNN-T speech recognition model bundle for the
CopperASR runtime. The model is configured for Russian speech recognition and
16 kHz audio input.

The repository uses a minimal normalized layout: three ONNX graphs, a token
mapping, and a runtime manifest. CopperASR can load this bundle directly without
additional conversion steps.

## Files

- `encoder.onnx`
- `decoder.onnx`
- `joint.onnx`
- `tokens.txt`
- `model_manifest.json`

## Runtime Contract

| File | Purpose |
| --- | --- |
| `encoder.onnx` | RNN-T acoustic encoder |
| `decoder.onnx` | RNN-T prediction network |
| `joint.onnx` | RNN-T joint network |
| `tokens.txt` | Token id mapping with `<blk>` at id `1024` |
| `model_manifest.json` | Runtime contract and checksums |

## Usage with CopperASR

```python
from copper_asr import CopperASR

asr = CopperASR(model_source="copperside-gigaam-v3-e2e-rnnt")
result = asr.transcribe("audio.wav")
print(result.full_text)
```

For a local checkout:

```python
asr = CopperASR(
    model_source="copperside-gigaam-v3-e2e-rnnt",
    model_path="path/to/CoppersideASR",
)
```

## Model Details

The model uses an RNN-T layout with separate encoder, decoder, and joint ONNX
graphs. Tokenizer settings, feature extraction parameters, decoder limits, and
file checksums are recorded in `model_manifest.json`.