ZipVoice-Distill GGUF for audio.cpp
Ready-to-run GGUF packages of ZipVoice-Distill (k2-fsa), a zero-shot voice-cloning TTS with a TTSZipformer flow-matching backbone and a Vocos 24 kHz vocoder, packaged for the audio.cpp engine.
Each file is fully self-contained: flow-matching model + bundled Vocos vocoder +
embedded text-frontend sidecars (tokens.txt, model.json, zh tables, jieba
dictionaries). Output: 24 kHz mono WAV.
Files
| File | Precision | Size | Recommended |
|---|---|---|---|
zipvoice-distill-q8_0.gguf | Q8_0 (weight-only) | 158 MB | ✅ default |
zipvoice-distill-orig.gguf | F32 (original) | 553 MB | bit-exact reference |
Q8_0 is the recommended download. Weights are stored as int8 blocks (one F16 scale per 32 weights); activations stay F32 and the kernels dequantize each block on the fly — no full-model dequantization at load, no int8 arithmetic.
Q8_0 vs F32 (Apple M3, Metal, 34 s utterance)
| Metric | F32 | Q8_0 |
|---|---|---|
| File size | 553 MB | 158 MB (3.5× smaller) |
| Runtime memory | 1310 MB | 971 MB (−26%) |
| RTF (lower is better) | 0.217 | 0.204 |
| Log-mel cosine vs F32 | 1.0000 | 0.9987 (male) / 0.9997 (female) |
| Velocity field cosine vs PyTorch | 1.0000 | 0.9993 |
Quality is transparent to the ear; 8-step distilled sampling, guidance scale 3, time shift 0.5. Lower-bit variants (q6_k … q4_0) are deliberately not provided — they trade audible quality for size on this model.
Usage
Install with the audio.cpp model manager
# from an audio.cpp checkout
python3 tools/model_manager_v2.py install zipvoice_distill_q8_0
# → models/ZipVoice-Distill-GGUF/zipvoice-distill-q8_0.gguf
Direct synthesis (CLI)
audiocpp_cli --task clon --family zipvoice \
--model models/ZipVoice-Distill-GGUF/zipvoice-distill-q8_0.gguf \
--voice-ref prompt.wav --reference-text "Reference transcript." \
--text "你好，这是ZipVoice，支持中文和English两种语言。" --out out.wav
--reference-text must transcribe --voice-ref. Chinese, English and mixed
input are handled by the built-in Emilia frontend (jieba + pypinyin tables +
espeak-ng; the English path needs the espeak-ng dylib shown above). Python API,
server and streaming options are documented in
docs/community_models/zipvoice.md.
Provenance & conversion
- Upstream weights: k2-fsa/ZipVoice
(
zipvoice_distill, Apache-2.0) - Conversion:
tools/community_models/convert_zipvoice.py(torch → safetensors → GGUF withmodel.*/vocos.*namespaces) - Quantization:
audiocpp_gguf --type q8_0(weight-only, per-block scales) - Parity: text encoder (per-layer taps), flow decoder, Vocos and full synthesis verified against the PyTorch reference; see the zipvoice doc in audio.cpp for the full report
License
Apache-2.0, inherited from the upstream ZipVoice release.