---

name: moss-tts description: "Route MOSS-TTS family speech, voice-agent, sound-effect, llama.cpp, streaming, and fine-tuning workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

MOSS-TTS

Use this repo skill when a task names MOSS-TTS, OpenMOSS/MOSS-TTS, MOSS-TTS-v1.5, MOSS-TTSD, MOSS-VoiceGenerator, MOSS-TTS-Realtime, MOSS-TTS-Local-Transformer-v1.5, MOSS-SoundEffect, MossTTSDelay, MossTTSLocal, moss-tts-llama-cpp, GGUF/ONNX/TensorRT audio-tokenizer inference, or MOSS-TTS fine-tuning JSONL.

This skill teaches how to operate the model-family workflows. It does not execute long model downloads, generation, training, service startup, benchmark runs, TensorRT builds, or GPU performance checks by default.

First route by task

User intent	Read next
Hugging Face AutoProcessor/AutoModel generation, voice cloning, duration control, language tags, Pinyin/IPA, MOSS-TTSD, MOSS-VoiceGenerator, SoundEffect v1, or Gradio-style model-family demos	sub-skills/hf-family-workflows/SKILL.md
Torch-free, low-memory, GGUF, ONNX, TensorRT, moss-tts-llama-cpp, config YAML, bridge build, weight conversion, or batch eval	sub-skills/llama-cpp-backend/SKILL.md
MOSS-TTS-Local-Transformer-v1.5 batch or realtime streaming decode, 48 kHz stereo, codec v2, browser streaming app, or token/frame estimates	sub-skills/local-v15-streaming/SKILL.md
MOSS-TTS-Realtime voice-agent sessions, streaming text deltas, FastAPI service, client payloads, KV-cache reuse, or prompt-audio context	sub-skills/realtime-voice-agent/SKILL.md
Fine-tuning manifests, JSONL schema, audio_codes, preprocessing, accelerate, FSDP, ZeRO-3, sharded outputs, or post-training smoke plans	sub-skills/finetuning-data-prep/SKILL.md
MOSS-SoundEffect v2 DiT/DAC/Qwen3 pipeline, separate environment, MossSoundEffectPipeline, sound-effect fine-tuning metadata, or export to HF format	sub-skills/soundeffect-v2/SKILL.md

If the user only says “generate speech with MOSS-TTS,” start with hf-family-workflows. If the user says “no PyTorch,” “GGUF,” “low memory,” or “llama.cpp,” route to llama-cpp-backend. If the user says “Realtime” with voice-agent turns, route to realtime-voice-agent, not Local v1.5 streaming.

Read root references when needed

• references/model-family-overview.md — model/checkpoint/architecture map, audio-format differences, and scenario boundaries.
• references/installation-profiles.md — dependency profiles, extras, Python/CUDA/FFmpeg notes, and the package exposure caveat.
• references/troubleshooting.md — cross-cutting install/import, backend, model download, audio, and packaging failures before routing deeper.
• references/repo-provenance.md — source snapshot and refresh baseline.
• references/repo-routing-metadata.json — structured router metadata used by managed repo-skill importers.
• scripts/check_moss_tts_environment.py — safe diagnostic helper for metadata/import/optional dependency checks; run it before blaming model code.

Installation triage

Use a clean Python 3.12 environment for most MOSS-TTS workflows. Choose the smallest profile that matches the task:

• HF generation / Gradio demos: install the top-level package with torch-runtime; choose a PyTorch wheel index matching the host backend. FFmpeg is required by torchcodec audio I/O.
• FlashAttention: add flash-attn only when CUDA, GPU architecture, dtype, and build memory support it. Otherwise use SDPA on CUDA or eager on CPU.
• llama.cpp backend: use llama-cpp profiles instead of the full torch runtime; add ONNX, TensorRT, or Torch heads only when that backend is selected.
• Fine-tuning: add finetune; add finetune-deepspeed only for ZeRO-3.
• SoundEffect v2: use its separate package/environment; it pins a different NumPy/Transformers/Diffusers/Torch stack and should not be mixed with the root MOSS-TTS runtime.

Minimal diagnostic command from this root skill directory:

python scripts/check_moss_tts_environment.py --json

If metadata is installed but imports such as moss_tts_delay fail, read references/installation-profiles.md and sub-skills/hf-family-workflows/references/model-packaging.md; the current package metadata may require running from a source checkout, adding the checkout to PYTHONPATH, or using a packaging layout that explicitly exposes source packages.

Safety and verification boundaries

• Do not treat a CPU import as proof that CUDA generation, FlashAttention, TensorRT, or training works.
• Do not run long generation, model downloads, service listeners, training jobs, or benchmark batches unless the user explicitly authorizes time, network, disk, and hardware use.
• Prefer bundled validators and config inspectors for cheap checks before expensive runtime work.
• Keep model family and code family aligned: Delay, Local 1.0, Local v1.5, TTSD, VoiceGenerator, SoundEffect v1, Realtime, and SoundEffect v2 have different prompt/code/tokenizer assumptions.
• For any current repository checkout whose commit or public APIs differ from references/repo-provenance.md, run refresh-repo-skill before relying on this skill for exact flags or compatibility rules.