---
license: mit
tags:
  - voice-conversion
  - any-to-any
  - onnx
  - vconnx
language:
  - en
pipeline_tag: audio-to-audio
---

# vconnx-triaan-vc

Any-to-any voice conversion using TriAAN-VC (Triple Adaptive Attention Normalization,
ICASSP 2023), exported to ONNX for CPU-first inference via the
[vconnx](https://github.com/TigreGotico/vconnx) engine adapter.

## Model card

| Field | Value |
|---|---|
| Paper | [TriAAN-VC: Triple Adaptive Attention Normalization for Any-to-Any Voice Conversion](https://arxiv.org/abs/2303.09057) |
| Original code | [winddori2002/TriAAN-VC](https://github.com/winddori2002/TriAAN-VC) |
| License | MIT (original code and weights) |
| Sample rate | 16 000 Hz |
| Mel channels | 80 |
| FFT / hop / win | 400 / 160 / 400 samples |

## Architecture

Three ONNX components chained in sequence:

1. **CPC encoder** — 5-layer strided Conv1d (160× downsample) + LSTM; encodes speech
   to 256-dim content features at 100 Hz.  Derived from
   [facebookresearch/CPC\_audio](https://github.com/facebookresearch/CPC_audio).
2. **TriAAN-VC decoder** — SpeakerEncoder + ContentEncoder + BiGRU fusion +
   TriAAN decoder + PostNet; maps CPC features + log-F0 → 80-band mel spectrogram.
3. **ParallelWaveGAN vocoder** — WaveNet-style neural vocoder (MIT, kan-bayashi); mel
   → 16 kHz waveform.

## Files

| File | Size | Description |
|---|---|---|
| `cpc_encoder.onnx` | 7.0 MB | CPC content encoder (FP32) |
| `cpc_encoder_q8.onnx` | 1.8 MB | CPC encoder INT8 dynamic quantization |
| `triaan_vc.onnx` | 266.3 MB | TriAAN-VC decoder (FP32) |
| `triaan_vc_q8.onnx` | 76.3 MB | TriAAN-VC decoder INT8 dynamic quantization |
| `pwg_vocoder.onnx` | 7.0 MB | ParallelWaveGAN vocoder (FP32) |
| `pwg_vocoder_q8.onnx` | 2.0 MB | PWG vocoder INT8 dynamic quantization |
| `config.json` | — | vconnx component manifest |
| `PROVENANCE.md` | — | Upstream provenance + license text |

## Parity (PyTorch vs ONNX Runtime, random input, T=50 frames)

| Component | max\_abs Δ | mean\_abs Δ | Verdict |
|---|---|---|---|
| CPC encoder | 1.06e-05 | 1.46e-07 | PASS |
| TriAAN-VC | 3.76e-06 | 6.39e-07 | PASS |
| ParallelWaveGAN | 4.39e-05 | 1.19e-06 | PASS |

All components pass with max\_abs < 5e-05 (well within perceptual threshold).

## Usage

```python
from vconnx import VoiceCloner

vc = VoiceCloner(engine="triaan")
output_wav = vc.clone_voice(source_wav, target_wav)
```

Install:

```bash
pip install "vconnx[triaan]"
```

## Provenance

Weights converted from the GitHub release v1.0 artifacts:

- `cpc.pt` — CPC encoder checkpoint (facebookresearch/CPC\_audio)
- `model-cpc-split.pth` — TriAAN-VC decoder checkpoint
- `vocoder.pkl` — ParallelWaveGAN vocoder checkpoint (originally from VQMIVC)

See `PROVENANCE.md` for full license text.