phoonnx-styletts2

Mirror of StyleTTS2-family checkpoints exported to ONNX for phoonnx, OpenVoiceOS's offline TTS library. StyleTTS2 generates speech conditioned on a style vector that encodes prosody and timbre from a reference clip, which is what enables zero-shot cloning; phoonnx also treats the Kokoro checkpoints (a StyleTTS2 descendant, "misaki"-phonemized) as styletts2-engine voices since they share the same style-vector conditioning contract. See docs/cloning.md for the styletts2_style d-vector mechanism.

This repository holds the raw ONNX weights, config.json files, and (for cloning-capable voices) .bin style vectors consumed by phoonnx's voice_index (styletts2.json); it is not meant to be browsed for voices directly — use phoonnx's voice manager, which resolves voice IDs to the files here.

Contents

Directory	Voices	Language	Upstream
ddatt-en-styletts2/	ddatt/en-styletts2	en-US	community StyleTTS2 checkpoint by ddatt; original training source could not be located
bsc-es-styletts2/	bsc/es-styletts2 (zero-shot) + 6 named speakers bsc/es-cml<id>	es	BSC-LT/styletts2-spanish-multispeaker (Barcelona Supercomputing Center)
bsc-ca-styletts2/	bsc/ca-styletts2 (zero-shot) + 11 named speakers bsc/ca-<name>	ca	BSC-LT/styletts2-catalan-multispeaker (Barcelona Supercomputing Center)
hitz-eu-styletts2/	one speaker each: antton, maider	eu	HiTZ/StyleTTS2-eu (HiTZ Center, University of the Basque Country)
hitz-eu-emo-styletts2/	antton/maider x happy/neutral/sad	eu	HiTZ/StyleTTS2-eu_emo (HiTZ Center)
proxectonos-gl-celtia/	proxectonos/celtia-styletts2	gl	proxectonos/Nos_StyleTTS2-Celtia-GL (Proxecto Nós, technical development by Gradiant)
proxectonos-gl-brais/	proxectonos/brais-styletts2	gl	proxectonos/Nos_StyleTTS2-Brais-GL (Proxecto Nós, technical development by Gradiant)
kokoro-82m/	kokoro/* (English voices, plus per-language configs for es/fr/hi/it/ja/pt/zh)	en-US + others	hexgrad/Kokoro-82M; style vectors fetched at load time from onnx-community/Kokoro-82M-v1.0-ONNX
kokoro-v0.19/	Kokoro v0.19 voices	en-US	hexgrad/kLegacy (Kokoro v0.19)
kokoro-v1.1-zh/	Kokoro v1.1 zh/en voices	zh, en-US	hexgrad/Kokoro-82M-v1.1-zh

Usage

import wave
from phoonnx.model_manager import TTSModelManager

manager = TTSModelManager()
manager.load()
manager.merge_default_voices()

voice = manager.voices["kokoro/af_heart"].load()
with wave.open("out.wav", "wb") as wav_file:
    voice.synthesize_wav("Hello, this is a test.", wav_file)

License

license: other because the checkpoints in this repository come from different labs with different licenses. phoonnx itself is Apache-2.0.

Family	Voices	License
Kokoro (all versions)	kokoro/*	Apache-2.0
BSC Spanish StyleTTS2	bsc/es-styletts2, bsc/es-cml*	Apache-2.0
BSC Catalan StyleTTS2	bsc/ca-styletts2, bsc/ca-*	GPL-3.0
ProxectoNós Galician StyleTTS2	proxectonos/celtia-styletts2, proxectonos/brais-styletts2	Apache-2.0
HiTZ Basque StyleTTS2 (base + emotion)	hitz-eu-styletts2/*, hitz-eu-emo-styletts2/*	Apache-2.0
ddatt/en-styletts2	ddatt-en-styletts2/	not determined — original upstream repo could not be located; treat as all-rights-reserved until clarified

The two BSC checkpoints are multispeaker. Neither upstream repository ships reference audio, so bsc/es-styletts2 and bsc/ca-styletts2 need a reference clip on every call. Alongside them each directory now carries one <speaker>.bin style blob per named speaker, so the same checkpoint also serves ordinary preset voices:

• bsc-ca-styletts2/ — 11 Festcat speakers: bet eli eva jan mar ona pau pep pol teo uri.
• bsc-es-styletts2/ — the 6 CML-TTS Spanish speakers that hold the corpus: 3946 8882 9972 10246 11797 12367. CML-TTS names speakers by number only.

Each blob is 256 float32 values (ref_p ++ ref_s), the mean of the style encoder's output over four reference clips of that speaker taken from the model's own training corpus. A named speaker still clones when you pass a reference clip. See docs/bsc_multispeaker.md.

The Galician voices are single-speaker: each ships a style.bin default style alongside the style_encoder.onnx used for cloning. They are phonemized with Cotovía notation (not IPA) — see docs/galician.md.

Note the BSC Catalan checkpoint is GPL-3.0-licensed by its authors; this only governs the model weights themselves, not phoonnx (Apache-2.0), which merely loads them.