MOSS-TTS-Realtime-GGUF

MOSS-TTS-Realtime GGUF

End-to-end GGUF conversion of OpenMOSS-Team/MOSS-TTS-Realtime, runnable as a backbone + codec_lm + codec stack via stock llama.cpp and codec.cpp.

The model is a Qwen3-2B language backbone + 4-layer MossTTSRealtimeLocalTransformer depth decoder + 17-channel emission (cb-0 = a text token sampled from the Qwen3 backbone's lm_head; cb-1..16 = 16 RVQ audio codebooks of 1027 entries each). codec.cpp's residual_depth_ar codec_lm runtime handles the depth decoder + audio embed tables + per-AR-step state machine; the backbone runs in llama.cpp as a stock qwen3 arch.

This replaces the earlier codec-only release: the LLM-part is now wired through codec.cpp's codec_lm infrastructure (see tts_remaining_plan.md in codec.cpp).

Files

Backbone (Qwen3 language_model, stock qwen3 arch — 28 layers, hidden 2048, vocab 151936)

moss-tts-realtime-<quant>.gguf — produced by codec.cpp's convert-backbone-to-gguf.py prep_moss_tts_realtime, which unwraps language_model.* into the standard Qwen3 layout. tie_word_embeddings=true is inferred from the absence of a standalone lm_head tensor.

Codec + codec_lm (MOSS-Audio-Tokenizer 16 RVQ × 1027 codebooks + residual_depth_ar adaptor)

codec[-<quant>].gguf — full MOSS-Audio-Tokenizer (1.6B, 24 kHz mono) bundled with the codec_lm adaptor: audio embed tables for the 16 RVQ heads, the 17th text_embd (~308 MB at F16), 4-layer depth decoder, and 16 codebooks_head slices. Produced by codec.cpp's convert-to-gguf.py --model-type moss_audio --lm-source OpenMOSS-Team/MOSS-TTS-Realtime.

Inference shape (text-modality codec_lm AR)

backbone (Qwen3, embeddings=true) hidden state h
    → caller samples text token via llama_get_logits_ith → text_tok
    → codec_lm_state_set_text_context(state, text_tok)
    → codec_lm_step_begin(state, h)
    → for cb in 0..16: codec_lm_step_logits → sample → codec_lm_step_push_code
    → codec_lm_step_finish → codes[17]
    → codec_lm_compose_audio_embd(codes) → next-step embedding
    → feed to backbone via b.embd; loop until codes[0] == EOS_text

After generation, slice cb-0 out of the (T × 17) code matrix (it's a text token, not a codec value) and feed cb-1..16 into codec_decode to get 24 kHz mono PCM.

Sources

• Upstream model: OpenMOSS-Team/MOSS-TTS-Realtime
• Audio codec: OpenMOSS-Team/MOSS-Audio-Tokenizer
• Conversion tooling: mybigday/codec.cpp
• Inference runtime: mybigday/llama.rn (codec_lm AR text-modality path in cpp/rn-tts.cpp)