---
license: cc-by-nc-sa-4.0
tags:
  - voice-conversion
  - onnx
  - bicodec
  - sparktts
  - vconnx
language:
  - en
---

# vconnx-bicodec

ONNX artifacts for the `bicodec` engine of [vconnx](https://github.com/TigreGoticoLda/vconnx) —
a pure-ONNX multi-engine voice-cloning library.

BiCodec (SparkAudio/Spark-TTS, 2025) achieves zero-shot any-to-any voice conversion via explicit
factorization of speech into two complementary token streams:

- **Semantic tokens** (content): Wav2Vec2-XLSR-53 (hidden layers 11, 14, 16 averaged) →
  convolutional encoder → FactorizedVQ → (1, T) int64
- **Global tokens** (speaker): 128-bin Slaney mel spectrogram → ECAPA-TDNN + Perceiver +
  FSQ → (1, 1, 32) int32 (32 fixed tokens per utterance)

Voice conversion: **source semantic tokens + reference global tokens → decoder → waveform**.
No auto-regressive language model; single forward pass per chunk.

## License

**Upstream weights: CC BY-NC-SA 4.0 — non-commercial use only.**

These ONNX artifacts are derived from [SparkAudio/Spark-TTS-0.5B](https://huggingface.co/SparkAudio/Spark-TTS-0.5B)
weights which are licensed under the
[Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License](https://creativecommons.org/licenses/by-nc-sa/4.0/).

- You may **not** use these artifacts for commercial purposes.
- You must give appropriate credit to SparkAudio.
- Derivative works must carry the same CC BY-NC-SA 4.0 license.

Upstream code (SparkAudio/Spark-TTS): Apache-2.0.

## Usage

```python
from vconnx import VoiceCloner

cloner = VoiceCloner(engine="bicodec")
out = cloner.clone_voice("source.wav", "reference.wav", "out.wav")
print(cloner.sample_rate)   # 16000
```

Install: `pip install vconnx`

## Components

| File | Description | Size (fp32) | Size (INT8) |
|---|---|---|---|
| `wav2vec2_encoder.onnx` | Wav2Vec2-XLSR-53 encoder (layers 11/14/16 avg) | 819 MB | 206 MB |
| `semantic_encoder.onnx` | Conv encoder + FactorizedVQ → semantic tokens | 116 MB | 30 MB |
| `global_encoder.onnx` | ECAPA-TDNN + Perceiver + FSQ → global tokens | 22 MB | 6 MB |
| `decoder.onnx` | Token decoder → waveform | 368 MB | 158 MB |
| `mel_filterbank.npy` | 128-bin Slaney mel filterbank (numpy, 128×513) | 256 KB | — |
| `mel_config.json` | STFT parameters (n_fft=1024, hop=320, win=640) | ~1 KB | — |

Output sample rate: **16 kHz**.

## Parity vs PyTorch

| Component | Metric | Value | Result |
|---|---|---|---|
| wav2vec2_encoder | max_abs | 6.71e-4 | PASS |
| semantic_encoder | exact int match | True | PASS |
| global_encoder | exact int match | True | PASS |
| mel numpy vs torchaudio | max_abs | 0.00e+0 | PASS |
| decoder | max_abs | 2.53e-6 | PASS |

## References

- [Spark-TTS paper](https://arxiv.org/abs/2503.01710)
- [SparkAudio/Spark-TTS](https://github.com/SparkAudio/Spark-TTS)
- [SparkAudio/Spark-TTS-0.5B](https://huggingface.co/SparkAudio/Spark-TTS-0.5B) (upstream weights)
- [vconnx](https://github.com/TigreGoticoLda/vconnx) (inference library)