cstr/moss-tts-v1.5-GGUF

MOSS-TTS-v1.5 — GGUF (for CrispASR)

GGUF conversion of OpenMOSS-Team/MOSS-TTS-v1.5 (MossTTSDelay: a Qwen3-8B backbone emitting 32 RVQ audio codebooks under a delay pattern) + its MOSS-Audio-Tokenizer 1.6B transformer codec, for the CrispASR moss-tts backend.

Files

File	What	Size
moss-tts-v1.5-q4_k.gguf	Q4_K backbone (default; audio tables kept F16)	~7 GB
moss-tts-v1.5-f16.gguf	F16 backbone (needs >20 GB VRAM or CPU; for re-quant)	~17 GB
moss-tts-v1.5-codec.gguf	F16 transformer codec companion	~3.5 GB

Use

crispasr --backend moss-tts -m moss-tts-v1.5-q4_k.gguf \
         --codec-model moss-tts-v1.5-codec.gguf \
         --tts "Hello world." --tts-output out.wav
# or: crispasr --backend moss-tts -m auto --auto-download --tts "..."

Validated on CUDA (P100) by decoded round-trip (synthesize → ASR): the Q4_K backbone produces intelligible, accurate speech end-to-end.

License

Apache-2.0, inherited from the base MOSS-TTS-v1.5 + MOSS-Audio-Tokenizer models (OpenMOSS-Team). This repo redistributes derived GGUF weights under the same terms.