Lyric Slice By Time

The reason you aligned a whole song in the first place is usually this: you have a time window - a clip that starts at 29.5 seconds and runs 7 seconds - and you want the exact lines being sung in it. LyricSliceByTime is that lookup. It takes the alignment JSON from LyricForcedAlign plus a start and a duration, and hands back the ground-truth lines covering that window. Because the answer is chosen by time, a repeated chorus is never ambiguous: the same lyric line at second 29 and at second 48 comes back as two separate, correctly-timed entries.

Inputs that matter

• alignment (STRING): the JSON output of LyricForcedAlign. Nothing else works here - this node doesn't align, it queries an existing alignment.
• start_index (FLOAT, seconds): clip start. Negative counts back from the end, the same convention as TrimAudioDuration - -30 starts 30 seconds before the track ends.
• duration (FLOAT, seconds): how long the window is.
• mode: overlap (default) keeps a line if enough of it falls inside the window; contained keeps only lines that fit entirely inside.
• min_overlap (FLOAT, seconds, default 0.3): the one that saves you. In overlap mode a line counts only if at least this many seconds of it sit in the window. Without it, a line that starts 0.08 s before the window ends is "in" the clip while the audio holds barely a syllable of it. Set 0 to restore bare any-overlap.
• rebase_lrc (BOOLEAN, default on): retimes the LRC so timestamps are relative to the clip start. Leave it on unless you specifically want track-absolute times.

How it works

Plain window math. It computes [start, start + duration], then filters the alignment's lines by overlap or containment. Nothing clever, and that's the point - it's deterministic, instant, and reruns any time you nudge the knobs.

Outputs

• text: the lines for the window, joined. Wire into a caption node, a subtitle pass, SaveText.
• lrc: the same lines as a rebased LRC file starting at 00:00.
• info: a one-liner saying which window you queried and how many of the song's lines landed in it.

Where it fits

It pairs with TrimAudioDuration: slice the alignment for the same window you cut the audio with, and you've got clip-plus-correct-text as one consistent pair. That's the whole workflow for "give me the lyrics for this exact 7-second clip" - a one-wire query instead of a manual scroll through the song.

Install and gotchas

Same pack as LyricForcedAlign - ComfyUI Manager search "comfyui_lyric_align" or git clone https://github.com/bmad4ever/comfyui_lyric_align into custom_nodes, restart, and the shared dependencies (torchaudio + rapidfuzz) cover it. This node itself is pure string math, so it's the cheapest thing in the pack.

If a line you expected is missing, check the mode. contained is strict and drops any line that crosses a boundary - switch to overlap with the default 0.3 s overlap if you want edge lines included.