Luigi/openvoice2-melo8k-zh-gguf

OpenVoice2 · MeloTTS ZH (8 kHz) — GGUF for RapidSpeech.cpp

A GGUF (ggml) build of the 8 kHz vocoder-distilled MeloTTS Chinese voice, for inference with RapidSpeech.cpp (a ggml-based unified speech framework). It is the ggml sibling of the ONNX model at Luigi/vits-melo-tts-zh_en-8k — same voice, same 8 kHz output, different runtime.

This is the merged artifact: OpenVoice2 base (ZH) acoustic model + the melo8k 8 kHz student vocoder, packaged as one GGUF so RapidSpeech.cpp can synthesize end-to-end.

Why 8 kHz

Built for telephony AI (the iPhO 2.0 edge IP-PBX on Jetson Nano). Phones capture and play 8 kHz mono, so a vocoder that natively outputs 8 kHz removes the resample and ~halves vocoder compute vs a 44.1 kHz model — the cost of a vocoder is per-output-sample.

The 8 kHz student is a Vocos8k distillation (ConvNeXt @ 125 Hz frame rate + iSTFT head, ~3.7 M params, PESQ-NB ≈ 2.90, MCD ≈ 10.8) of the full MeloTTS vocoder — ~10× cheaper, same voice / prosody / code-switching.

Model details

Architecture (gguf general.architecture)	openvoice2
Voice	MeloTTS Chinese (ZH), zh-en code-switching
Output sample rate	8 kHz mono
Vocoder	melo8k — Vocos8k student (ConvNeXt + iSTFT, ~3.7 M params)
Format	GGUF (f32 merge), ggml runtime
Size	~156 MB
Runtime	RapidSpeech.cpp (CPU + CUDA; gen1 Jetson sm_53 supported)

Usage (RapidSpeech.cpp)

# CPU
rs-tts-offline -m openvoice2-melo8k-zh.gguf \
  -t "您好，請問您要找哪一位同仁？" --lang Chinese -o out.wav --gpu false

# CUDA (e.g. Jetson Nano gen1)
rs-tts-offline -m openvoice2-melo8k-zh.gguf \
  -t "您好，請稍候，我幫您轉接。" --lang Chinese -o out.wav --gpu true

Language flag matters. Pass --lang Chinese — this voice is ZH-only. (In the ONNX/sherpa-onnx sibling the equivalent is --sid 1; the default English/speaker-0 path feeds the wrong embedding and produces silence.)

Output is 8 kHz mono PCM WAV — feed directly to a G.711 telephony leg, no resample.

Relationship to the ONNX model

	This repo (GGUF)	vits-melo-tts-zh_en-8k (ONNX)
Runtime	RapidSpeech.cpp (ggml)	sherpa-onnx / ONNX Runtime
Role in iPhO 2.0	ggml backup framework	ONNX primary framework
Output	8 kHz mono	8 kHz mono
Speaker selector	--lang Chinese	--sid 1

On CPU the ONNX/sherpa path is faster; RapidSpeech.cpp's advantage is the ggml CUDA backend on Jetson. Both produce the same 8 kHz voice.

Provenance

melo8k student distillation and the OpenVoice2 merge were produced locally for iPhO 2.0; the merge recipe (convert_melo8k_merge.py) lives in the RapidSpeech.cpp fork. Base voice © MeloTTS / MyShell OpenVoice2 (MIT). This packaging is released under MIT.