MOSS-TTS Speak

MOSSSpeak is the "give me a voice out of thin air" node. You hand it text, a language, and an instruction like "male, warm, elderly narrator", and out comes stereo speech - no reference clip, no API key, no account. It's the part of the MOSS-TTS 1.5 pack you reach for when you need narration on a deadline and nobody's voice is in the pipeline yet. Open TTS has gotten genuinely good at this job locally, and MOSS's angle is that you don't need to hunt down a sample to steal a voice from - you just describe one.

How the "no reference" trick works

MOSS was trained on a path where the reference is a literal "None" placeholder. When you don't supply audio, it invents a voice from language + instruction. That makes instruction your only voice-steering knob here, so spend ten seconds on it - "male, warm, elderly narrator" and "young female, cheerful, energetic" genuinely produce different characters. Leave it empty and MOSS falls back to whatever the training-data default was for that language.

Inputs that matter

• moss_model - the bundle from MOSS-TTS Load Model.
• text - what to say. Multiline.
• language - one of 31. It also nudges MOSS toward a language-typical base voice, so set it explicitly rather than trusting detection.
• instruction - the voice description above. The one real knob.
• audio_temperature - default 1.7 (MOSS's own). Lower = safer and flatter, higher = more expressive and noisier. If output sounds unhinged, drop toward 1.2–1.4 before touching anything else.
• target_tokens - duration hint in frames; 0 = let the model decide. 375 ≈ 30 s. Pair it with the pack's Estimate Tokens node so you're not guessing.
• seed - same seed + same inputs = identical output, which makes retakes predictable.

Outputs: audio (stereo at the loaded model's native rate - 48 kHz on the default 1.7B) straight into SaveAudio / PreviewAudio, and tokens_generated (frames, ÷12.5 = seconds), which you can wire into a Voice Continue's previous_tokens if this clip is segment one of a longer read.

Optional things you'll eventually touch

audio_repetition_penalty (1.0 = off) is the fix for the classic autoregressive-TTS failure modes - droning, tempo freeze, looping syllables. A mild 1.05–1.15 usually kills them without flattening prosody; above ~1.3 it starts mangling legitimate repeats like "nein, nein, nein". The text_* samplers are a separate story: MOSS v1.5 is dual-stream, so text_temperature, text_top_p, and text_top_k steer the alignment/pacing channel, not the acoustics. Lowering text_temperature a notch gives steadier pacing without flattening the voice.

Install & common issues

Same pack as the rest - Manager, search "MOSS-TTS 1.5", or:

cd ComfyUI/custom_nodes
git clone https://github.com/eehrich/ComfyUI-MOSS-TTS-1.5.git MOSS-TTS-ComfyUI

Restart, run the Load Model node once, and the weights stream down on first execution. Two things bite people:

• Empty or whitespace-only text used to hang for minutes (MOSS never emits EOS with nothing to say). It now fails fast with a clear error instead - that's a feature, not a bug.
• [pause 1.2s] markers are spoken literally. MOSS v1.5 has no pause-marker parser. For a real gap, generate two clips and splice a silence spacer between them, or use punctuation (,, --, ...) to shape natural breaks.

The bundled example workflow does the sensible thing: Speak to create a voice from nothing, then feed that clip into Voice Clone to narrate the rest in the same voice.