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

# vconnx-openvoice-v2

ONNX export of the **OpenVoice v2** tone-color converter
([myshell-ai/OpenVoice](https://github.com/myshell-ai/OpenVoice), MIT license)
for use with [vconnx](https://github.com/TigreGotico/vconnx) — a pure-ONNX
zero-shot voice conversion toolkit.

## Export details

Both components are exported from the upstream `SynthesizerTrn`
(`myshell-ai/OpenVoice`) with **strict state-dict loading** — no architecture
reconstruction.  The checkpoint loads with 0 missing keys and 0 unexpected keys.

| Component | File | Size |
|---|---|---|
| Reference encoder (FP32) | `tone_ref_encoder.onnx` | 3.1 MB |
| Reference encoder (INT8) | `tone_ref_encoder_q8.onnx` | 2.2 MB |
| Voice converter (FP32) | `tone_converter.onnx` | 122.1 MB |
| Voice converter (INT8) | `tone_converter_q8.onnx` | 38.9 MB |

## Architecture

| Sub-graph | ONNX inputs | ONNX output |
|---|---|---|
| `tone_ref_encoder.onnx` | `spec` `(B, T, 513)` float32 — linear STFT magnitude | `tone_embedding` `(B, 256)` |
| `tone_converter.onnx` | `spec` `(B, 513, T)`, `spec_lengths` `(B,)`, `src_g` `(B, 256, 1)`, `tgt_g` `(B, 256, 1)` | `audio` `(B, 1, samples)` float32 — raw waveform |

The converter includes the full VITS-style flow decoder **and** HiFi-GAN vocoder;
it outputs raw audio directly.  No separate vocoder step is needed at inference.

**Preprocessing:** linear magnitude spectrogram matching upstream
`spectrogram_torch` — Hann window, n_fft=1024, hop=256, win=1024, reflect-pad
384 on each side, `sqrt(Re² + Im² + 1e-6)`.  No log compression.

## Parity (upstream torch vs ONNX)

| Component | max_abs_delta | mean_abs_delta | Status |
|---|---|---|---|
| `tone_ref_encoder` | 8.64e-07 | 2.45e-07 | **PASS** |
| `tone_converter` (5 seeds) | 1.08e-02 (worst) | 1.26e-04 (avg) | **PASS** |

The converter `max_abs` divergence is due to float32 accumulation through 4
residual coupling blocks in the flow — the quality-relevant metric is
`mean_abs`, which passes at 1e-3.

## E2E sanity check

Converted a 2 s synthetic source (220 Hz harmonics) to a 330 Hz reference:

| Metric | Value |
|---|---|
| Output duration | 1.997 s (source 2.000 s, ratio 0.998) |
| RMS | 0.268 |
| Spectral flatness | 0.072 (tonal, not noise) |
| Sample rate | 22050 Hz |

## Usage

```python
from vconnx import VoiceCloner

cloner = VoiceCloner(engine="openvoice")
cloner.clone_voice("source.wav", "reference.wav", "output.wav")
```

Or with the low-level adapter:

```python
from vconnx.engines.openvoice import OpenVoiceV2Adapter

adapter = OpenVoiceV2Adapter(quantized=False)
adapter.clone_voice("source.wav", "reference.wav", "output.wav")
```

Install with: `pip install vconnx[openvoice]`

## Provenance

- Upstream weights: [myshell-ai/OpenVoiceV2](https://huggingface.co/myshell-ai/OpenVoiceV2)
- Upstream source: [myshell-ai/OpenVoice](https://github.com/myshell-ai/OpenVoice)
- License: MIT ("Starting from April 2024, both V2 and V1 are released under MIT
  License. Free for commercial use." — official README)
- Export method: legacy TorchScript ONNX exporter (`dynamo=False`), opset 14
  (new dynamo exporter fails on GRU)
- Strict load: 0 missing keys, 0 unexpected keys