lai-align

name	lai-align
description	Align existing captions to audio/video with word-level precision using the Lattice-1 model. Trigger when the user has both a media file AND a caption/transcript that need to be synchronized, or says "fix caption timing", "字幕对不上", "对齐字幕", "word-level timestamps", "karaoke timing", "timestamps are off". Do NOT trigger without existing text — use `/lai-transcribe` first.
allowed-tools	Read, Bash(lai:*)

LattifAI Forced Alignment

Produces word-level and segment-level timestamps with sub-frame accuracy. Requires an API key — see /lai-setup.

Basic Command

Positional args are input_media, input_caption, output_caption. Pick a <base> (the media stem or YouTube ID) and reuse it; outputs land in the current directory:

# <base> = podcast (from podcast.mp3)
lai alignment align podcast.mp3 podcast.srt podcast.aligned.json
# shortcut:
lai-align podcast.mp3 podcast.srt podcast.aligned.json
# or explicit key=value style (interchangeable, useful inside scripts):
lai alignment align --direct -Y \
    input_media=podcast.mp3 \
    input_caption=podcast.srt \
    output_caption=podcast.aligned.json \
    caption.input.split_sentence=false

Format is inferred from the output file extension (30+ formats — see /lai-caption). Add --direct -Y for non-interactive pipeline runs.

Output naming: prefer <base>.aligned.json for downstream pipelines (karaoke / translate / diarize all consume aligned JSON). Use <base>.srt etc. only when the alignment result is the final deliverable.

Common Options

Append as key=value pairs:

• caption.input.split_sentence=false (default in examples) — preserve the source caption's original cue boundaries. Set to true to re-segment into natural sentences via wtpsplit; useful when the source has mid-sentence cue breaks (typical for YouTube auto-captions) and you want clean segments for downstream karaoke / translation / summarization
• caption.render.word_level=true — keep per-word timestamps in the output (needed for karaoke; use a JSON or ASS output path)
• media.streaming_chunk_secs=300 — process audio >10 min with bounded memory (up to 20 h)
• alignment.device=auto — override auto device (cuda/mps/cpu)

Full list: lai alignment align --help.

Word-Level JSON Output

{
  "supervisions": [{
    "text": "Hello world",
    "start": 0.0,
    "end": 2.5,
    "words": [
      {"word": "Hello", "start": 0.0, "end": 0.5},
      {"word": "world", "start": 0.6, "end": 2.5}
    ]
  }]
}

Fan out to other formats (karaoke ASS, TextGrid, SRT, …) via /lai-caption.

Common Issues

Problem	Fix
API KEY verification error	Run lai auth trial or lai auth login — see /lai-setup
Timing looks wrong	Try caption.input.split_sentence=true or caption.input.normalize_text=true

Related Skills

• /lai-transcribe — generate a transcript first
• /lai-diarize — add speaker labels to aligned output
• /lai-caption — convert aligned JSON to any format
• /lai-youtube — YouTube URL → aligned captions in one command