---
tags:
- text-to-speech
- gguf
- ggml
- x-voice
library_name: ggml
---

# X-Voice GGUF

Self-contained GGUF bundles for the X-Voice GGML C++ runtime.

Runtime repository:

```text
https://github.com/bluryar/X-Voice.cpp
```

Upstream model source:

```text
https://huggingface.co/XRXRX/X-Voice
https://github.com/sunnyxrxrx/X-Voice
```

## Files

| file | approximate size | sha256 |
|---|---:|---|
| `x-voice-f32.gguf` | 1.6 GB | `077d8dec4d94ab4562ae31cab4c0e000a9bb63a831f9ee3a5da7c8e7587df347` |
| `x-voice-f16.gguf` | 1.1 GB | `82b8bf7604da1f64c4d98e8e6afc11b3afda57058627a01054e8fc50b44505b3` |
| `x-voice-q8_0.gguf` | 848 MB | `8cb3452ad7aa1730047572bd38f22a92f87256a7dc803124b93a955b1f8ae5a5` |
| `x-voice-q6_k.gguf` | 790 MB | `5c4eb825d890f69fc706c4e06b5dd5ddec8b4ff3311cde0fc12bfa43fbcc4cf9` |
| `x-voice-q4_k.gguf` | 729 MB | `776f05d88d7bc252298272159ab91176029f4238b7eb7902d982e83bce8746a5` |

The quantized files use a conservative X-Voice policy: large GGML `mul_mat`
matrix weights are converted, while conv, norm, bias, embedding, positional, and
small tensors remain in their source type.

## Quantization Benchmark

Single-run local benchmark on an RTX 4060 Ti with `NVIDIA_TF32_OVERRIDE=0`,
`--preset product`, `cfg_nonlayered`, 32 sampler steps, the default zh sample
text, and `/root/code/ggbond/models/test.wav`:

| model | size | wall | sampler | load | mel max abs vs f32 | mel mean abs vs f32 | wav max abs vs f32 | wav mean abs vs f32 |
|---|---:|---:|---:|---:|---:|---:|---:|---:|
| f32 | 1544.1 MiB | 10.74s | 9.222s | 0.515s | 0 | 0 | 0 | 0 |
| f16 | 1060.5 MiB | 8.89s | 7.316s | 0.513s | 1.40578 | 0.01232 | 0.89212 | 0.01086 |
| q8_0 | 847.4 MiB | 8.52s | 7.268s | 0.303s | 0.78026 | 0.02397 | 1.01920 | 0.01708 |
| q6_k | 790.0 MiB | 8.66s | 7.407s | 0.301s | 0.89031 | 0.04256 | 0.68326 | 0.02387 |
| q4_k | 728.8 MiB | 8.61s | 7.277s | 0.319s | 5.72519 | 0.28411 | 1.22644 | 0.07608 |

Interpretation: f16/q8_0/q6_k are the practical release candidates. q4_k is
size-first and should be auditioned carefully before use as a quality default.

## Quick Start

```bash
git clone --recursive https://github.com/bluryar/X-Voice.cpp
cd X-Voice.cpp
bash scripts/dev/build_xvoice_cuda.sh

NVIDIA_TF32_OVERRIDE=0 build-cuda/x-voice-cli \
  --model /path/to/x-voice-f32.gguf \
  --load-tensors \
  --synthesize \
  --text '发帖人（弟弟）在详细描述其五口之家的现状后，寻求处理家庭问题的建议，并提出了自己的初步计划。' \
  --text-kind plain \
  --language zh \
  --ref-wav /path/to/reference.wav \
  --preset product \
  --output-wav /tmp/xvoice.wav \
  --metadata-json /tmp/xvoice.json \
  --progress \
  -b cuda -t 8
```

## License

The C++ runtime is Apache-2.0. These model artifacts are converted from upstream
X-Voice weights; please follow the upstream model license and terms for any
redistribution or commercial use.