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.

Provenance and EU AI Act Art. 53 note

• Upstream model: OpenMOSS-Team/MOSS-TTS-v1.5 — published by OpenMOSS-Team.
• Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository. No training-content summary was found on the upstream model card at the time of writing; that documentation gap is upstream's and is not filled here.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.