ZipVoice TTS assets

ZipVoice is a zero-shot voice-cloning text-to-speech model (k2-fsa/ZipVoice). Moonshine ships an ONNX Runtime port as an optional TTS engine selected with a zipvoice_ voice prefix (see core/moonshine-c-api.h).

Files in this directory

Runtime assets, resolved under g2p_root (or supplied in memory via the C API). .ort is preferred over .onnx when both are present.

Key	Role
`text_encoder.onnx` / `.ort`	Embeds phoneme tokens, computes duration/upsampling.
`fm_decoder.onnx` / `.ort`	Flow-matching velocity decoder (Euler ODE).
`vocoder.onnx` / `.ort`	Vocos mel -> waveform.
`tokens.txt`	Phoneme/pinyin -> id lexicon (espeak IPA + pinyin).
`model.json`	Feature/architecture config (feat_dim, sample rate).

These weights are not committed to git in bulk; regenerate them from the upstream checkpoint (they are tracked with Git LFS when present). The Moonshine binary release tarballs do not contain them — they are downloaded/cached at runtime like the Kokoro/Piper assets.

Regenerating

Requires a local checkout of the ZipVoice repo with its dependencies installed:

bash

python3 scripts/export_zipvoice_model.py \
    --zipvoice-repo ~/projects/ZipVoice \
    --model-name zipvoice_distill

This runs the ZipVoice tooling (scripts/export_onnx.py, scripts/rewrite_swoosh.py, scripts/onnx_vocoder.py, scripts/check_onnx_parity.py) and converts the result to .ort.

Size vs quality

By default the smallest parity-passing set is shipped: mixed int8 acoustic models (per-channel int8 weights, with activation-limited output projections kept in fp32) plus the fp32 vocoder. The fm_decoder drops from ~455 MB (fp32) to ~119 MB (mixed int8) with no audible quality loss.

Pass --swoosh to instead ship the fp32 fm_decoder rewritten to use the ai.zipvoice custom ops (SwooshL/SwooshR/GluGate/DepthwiseConv1d/BiasNorm/Bypass). Those ops are compiled directly into libmoonshine (see core/moonshine-tts/src/zipvoice-custom-ops.cpp) — no shared library is loaded at runtime — so the custom-op graph runs everywhere, including iOS. The int8 graph does not reference the custom ops; they remain available for the fp32 path.

Distill vs full model

zipvoice_distill (default) uses 8 sampling steps / guidance 3.0; the full zipvoice model uses 16 steps / guidance 1.0. Inference is otherwise identical (the classifier-free-guidance logic is baked into the exported graph), so both deploy through the same files here; select with the zipvoice_model option.

Reference voices

The built-in reference voices (zipvoice_american_female, zipvoice_indian_male, …) are VCTK clips compiled directly into libmoonshine (core/moonshine-tts/src/zipvoice-voices-data.cpp), regenerated with scripts/export_zipvoice_voices_for_cpp.py. Callers can also pass their own reference clip as in-memory PCM (zipvoice/prompt_audio).