cstr/confucius4-tts-GGUF

Confucius4-TTS — GGUF for CrispASR

GGUF conversions of netease-youdao/Confucius4-TTS (Apache-2.0) for the CrispASR native runtime.

Confucius4-TTS is a zero-shot voice-cloning TTS: a GPT-2 text→semantic stage (T2S, with ECAPA-TDNN speaker conditioning on w2v-BERT features), a flow-matching DiT+WaveNet semantic→acoustic stage (S2A, with CAMPPlus style conditioning and a reference-mel prompt), and a BigVGAN 22.05 kHz vocoder.

File	Stage	Quant	Size
confucius4-tts-t2s-f16.gguf	T2S	F16	1.27 GB
confucius4-tts-t2s-q8_0.gguf	T2S	Q8_0	685 MB
confucius4-tts-t2s-q4_k.gguf	T2S	Q4_K	376 MB
confucius4-tts-s2a-f16.gguf	S2A (+ CAMPPlus)	F16	213 MB
confucius4-tts-s2a-q8_0.gguf	S2A (+ CAMPPlus)	Q8_0	162 MB
confucius4-tts-s2a-q4_k.gguf	S2A (+ CAMPPlus)	Q4_K	135 MB
confucius4-tts-bigvgan-22k-f16.gguf	Vocoder	F16	214 MB

The T2S GGUFs carry the baked LlamaTokenizer vocab+merges; the S2A GGUFs carry the funasr/campplus style encoder under campplus.*. BigVGAN is nvidia/bigvgan_v2_22khz_80band_256x (MIT).

Usage:

crispasr --backend confucius4-tts -m confucius4-tts-t2s-q4_k.gguf \
  --codec-model confucius4-tts-s2a-q4_k.gguf \
  --tts "The quick brown fox jumps over the lazy dog." -l en \
  --voice ref.wav --i-have-rights --tts-output out.wav

Credits: NetEase Youdao (Confucius4-TTS, Apache-2.0); FunASR/ModelScope (CAMPPlus); NVIDIA (BigVGAN).