Nimaone/pocket-tts-farsi-v2-onnx

English | فارسی

pocket-tts-farsi-v2 — ONNX edition (no PyTorch)

An ONNX conversion of mehdi-hf/pocket-tts-farsi-v2 — Persian text-to-speech with voice cloning, fully offline on CPU, with no PyTorch dependency.

This package is the output of the persian_tts project. The weights are a pure format conversion of the original model — no weights were modified.

Why this version?

	ONNX (this package)	Original (PyTorch)
Dependencies	onnxruntime, numpy, scipy, soundfile, sentencepiece	torch, transformers, pocket-tts
Install size (Windows)	~200MB	~1.2GB
GPU required	No	No
Startup	Fast	Slower
Accuracy	waveform diff 7.5e-4 vs reference (inaudible)	reference
Speed (2.8s of audio, CPU)	3283ms	3423ms

Package contents

flow_lm_step.onnx (+.data)        unified FlowLM graph — three modes in one graph (voice/prompt/generation)
mimi_encoder.onnx (+.data)        reference voice -> latent
mimi_decoder_step_kv.onnx (+.data) latent -> audio (1920 samples per step)
g2p_encoder.onnx                  G2P: T5 encoder stage
g2p_decoder.onnx (+.data)         G2P: decoder (greedy loop runs host-side)
weights.npz                       host constants: LUT table, speaker_proj, normalize stats
decode_state_init.npz             initial decoder state template
manifest.json                     input/output spec for every graph

Usage

Download and run with the ONNX engine from the persian_tts project:

git clone https://github.com/nimaone/persian_tts
cd persian_tts

# download this package
pip install -U "huggingface_hub[cli]"
hf download Nimaone/pocket-tts-farsi-v2-onnx --local-dir model/onnx

# install the lightweight dependencies (~200MB, no torch)
python -m venv env
./env/Scripts/python -m pip install onnxruntime numpy scipy soundfile sentencepiece

# run
./env/Scripts/python scripts/tts_onnx.py "سلام، حال شما چطور است؟"

License and credits

Important — non-commercial use only: the original TTS model is released under CC-BY-NC-4.0, and this restriction carries over to any format conversion of it, including this ONNX package. Commercial use requires a license from the original author.

Component	Original model	Author	License
Persian TTS + voice cloning	mehdi-hf/pocket-tts-farsi-v2	mallahyari — training code/pipeline: mallahyari/pocket-tts	CC-BY-NC-4.0
Persian G2P (grapheme-to-phoneme)	mehdi-hf/Homo-GE2PE-Persian-HF	repackaging of MahtaFetrat/Homo-GE2PE-Persian by Elnaz Rahmati et al.	MIT
Base library and architecture	pocket-tts	Kyutai	MIT

This package contains a derivative of a CC-BY-NC-4.0 work and, per the terms of that license, is redistributed under the same license with attribution to the original author. The G2P portion remains under MIT.

All model credit belongs to the original authors — this repository only provides the format conversion (PyTorch -> ONNX) and a lightweight runtime.