MOSS

MOSS

Separate speech and sound-effect controls with no installed workers. The manual documents the speech decoder limitation. Captured September 25, 2026.

MOSS is two local engines on one page:

• MOSS-TTS Local v1.5 — text to speech
• MOSS-SoundEffect v2.0 — text to sound effects

They are separate model families, separate workers, separate VRAM. Voice & TTS (Soon) links here, but fresh MOSS-TTS generation is currently blocked on the tested build. MOSS-SoundEffect remains a qualified sound-effects path.

Open Audio → MOSS.

Refresh reloads install and worker state. Open Models jumps to Model library. Each engine also has a Server button to Runtime.

Shared operator rules

1. Install the engine (downloads model, tokenizer/source, isolated runtime).
2. Start worker on an explicit GPU with enough free VRAM. MOSS is large; on the tested machine it did not fit a 12 GB auxiliary card and loaded on a 24 GB card. Generic autospawn is not capacity-aware — pick the device on Runtime if the one-click start lands on the wrong GPU.
3. Generate. The GUI posts for a WAV blob and plays it in an <audio> element. Successful API responses also persist under Omni output (kind=omni) with X-Omni-Output-* headers.
4. Unload/kill the worker when you need the GPU for Comfy or Music.

Do not run MOSS-TTS and ACE XL at the same time unless both APIs and host RAM prove they fit. Schedule them sequentially by default.

MOSS-TTS

Status row: installed badge, worker ready/starting/none, Install or Start worker.

Banners:

• not installed — install to download the model, paired audio tokenizer, source checkout, and isolated runtime
• installed but no worker — start a worker before generating

Fields:

• Speech text
• Language (default English)
• Seed
• Max frames (1–7500, default 2048)
• Temperature (0.1–3.0, default 1.7)

Generate speech stays disabled until the worker is ready. The player appears under the button.

API equivalent: POST /api/tts/moss_tts with response_format: wav. WAV is the dependable path and is required if the gateway restimes speed. speed on the API ranges 0.25–4.0. Other format names exist (mp3, ogg, flac, opus); WAV is what this tab uses.

The output URL can be fed to ACE-Step Vocal→BGM, A2A, or Cover as init_audio_url.

TTS caution

Earlier 48 kHz speech runs succeeded on a 3090, but a fresh load followed by inference failed in the upstream decoder with CUDA driver error: unknown error and exhausted WSL dxgresource descriptors. Omni retires the worker after a worker-side 5xx so poisoned CUDA state is not reused. Do not blindly retry or treat a ready worker as proof of working speech. Use a previously persisted WAV or a recorded source until a fresh run and cleanup are qualified. Voice-cloning breadth has not been characterized.

This tab is not MiniCPM TTS and not Qwen Talker.

MOSS-SoundEffect

Same install/start pattern for moss_sfx.

Fields:

• SFX prompt — describe one event ("crisp rhythmic click-clack of fast typing on a mechanical keyboard"), not a song
• Seconds 0.1–30 (default 10)
• Steps 1–200 (default 100)
• CFG 0–20 (default 4)
• Sigma shift 0–20 (default 5)
• Seed

Generate SFX posts /api/moss/sfx and plays the 48 kHz WAV.

VRAM estimate is about 10–12 GB; still prefer the larger GPU when the smaller card is busy. Subjective event quality requires listening in Media library. Hold the semantic event constant if you A/B against Audio Lab effects; the engines do not share a latent space or seed meaning.

ACE-Step Text2Samples is not a third dependable SFX path (unreleased adapter).

After generation

1. Confirm the file in Media library → Omni (TTS / STT).
2. Pin keepers.
3. Kill the exact moss_tts / moss_sfx worker on Runtime.
4. Recheck device VRAM before starting Comfy.

Related pages

• Voice & TTS — why the speech hub is still Soon
• Audio Lab — alternative effects
• Music — Vocal→BGM handoff
• Runtime