MOSS-TTS Voice Continue

MOSSVoiceContinue is the "don't make me re-clone it" node. You've generated a clip in some voice - via Speak, Voice Clone, or an earlier Continue - and you want the same voice to keep talking through the next paragraph. This node extends a prior MOSS clip with new text, inheriting the voice straight from the audio. No new reference clip, no re-clone, no seam. It's how you narrate a multi-paragraph passage without generating the whole thing at once.

The mechanism that bites people

MOSS is a prefix-continuation model. It doesn't just take audio and "say more" - it needs to know where it is in the script. So the node concatenates previous_text + " " + text into the full script, conditions on the prior audio, and generates audio for the new part. Which means previous_text must be the exact text that produced previous_audio - word-for-word, punctuation included. Get it wrong and you get Kauderwelsch, because the model can't align its position in the script. It's the single most common mistake, and the tooltip spells it out for a reason.

Inputs that matter

• moss_model - from the loader.
• previous_audio - the prior MOSS output.
• previous_text - the exact text behind that audio. Non-negotiable.
• text - the new lines, appended after previous_text.
• previous_tokens - the exact frame count of the prefix. Wire the upstream node's tokens_generated output here for a precise handoff; leave at 0 and the node measures from the audio duration (off by at most one frame).
• target_tokens - target length of the NEW segment only. The node adds the prefix internally because MOSS reads its hint as TOTAL (prefix + new) in continuation mode.
• head_trim_frames - default 1 (~80 ms) trimmed from the start of the new audio. MOSS's conv codec lets the last prefix frame bleed into the start of the continuation; this removes it. Set to 0 if you hear a chopped syllable, higher if the bleed persists.

Outputs: audio (new segment only, for per-segment QC - you hear just the delta), tokens_generated (frames of the new segment), full_audio (prefix + new, concatenated and resampled to the model's native rate), and full_tokens (cumulative frame count). The chain pattern:

seg N:   Voice Continue -> audio, full_audio, full_tokens
seg N+1: previous_audio  <- seg N.full_audio
         previous_tokens <- seg N.full_tokens

The VRAM trap

Chaining segment N+1 with the whole cumulative history is the classic failure mode: VRAM drifts up about 1 GB per 20 s of history and eventually OOMs mid-scene. The fix the author documents: pass only the last segment's audio as previous_audio (and its full_tokens as previous_tokens) rather than the whole concatenation. Prefix-continuation semantics still hold - MOSS aligns the prefix at the end of previous_text inside the full script - just with a shorter history. Watch nvidia-smi during a long scene to spot the drift early.

Install & troubleshooting

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 and load the model once first. A few notes from the README and source: an empty follow-up text is technically allowed but MOSS closes out almost immediately, so supply real lines. And like every generate node in the pack, a mild audio_repetition_penalty (1.05–1.15) is the cure if a segment starts droning or looping syllables. When target_tokens is set, target_overshoot_frames caps how far past it the model can run - a fuse against hangs on pathological input.

The neat trick: because the voice lives in the audio, not a separate reference, you can chain an arbitrary number of segments and each one inherits the voice for free. That's the entire point - one voice, one long take, no seams.