Forced Alignment
Generate word-level timestamps by aligning a transcript to its corresponding audio.
Overview
The Forced Alignment node uses the ElevenLabs Forced Alignment API to match each word in a provided transcript to its exact position in the audio. The output is structured JSON data containing word-level start and end timestamps. This is a data-producing node – it outputs timing information, not audio.
Configuration
| Field | Type | Default | Description |
|---|---|---|---|
| Transcript | string (textarea, required) |
"" |
The full text transcript of the audio input. Must accurately match what is spoken in the audio |
Inputs & Outputs
- Input:
in– audio file to align against the transcript - Output:
data– JSON array of word-level timestamps (not audio)
Output Format
The output is an array of AlignmentWord objects:
[
{ "word": "Hello", "start": 0.0, "end": 0.35 },
{ "word": "world", "start": 0.38, "end": 0.72 }
]
| Field | Type | Description |
|——-|——|————-|
| word | string | The aligned word |
| start | number | Start time in seconds |
| end | number | End time in seconds |
Best Practices
- The transcript must closely match the actual spoken content for accurate alignment. Mismatches between the transcript and audio will produce unreliable timestamps.
- Clean up the transcript before alignment – remove filler words, false starts, or non-speech annotations unless they are actually spoken in the audio.
- Use the Voice Extractor node upstream if the audio has significant background noise, as clean audio produces better alignment accuracy.
- For long audio files, ensure the transcript covers the full duration. Partial transcripts will only align the covered portion.
Common Use Cases
- Generating precise word-level timestamps for subtitle/caption creation
- Building karaoke-style word highlighting synchronized to audio
- Creating timed animations or motion graphics that react to specific words
- Synchronizing visual elements to speech in video compositions
- Quality-checking TTS output timing against expected pacing
Tips
- This node outputs
data, notaudio. The output handle is"data", which means it connects to data-consuming nodes rather than audio-consuming ones. - The transcript should be plain text without formatting, timestamps, or speaker labels. Just the words as spoken.
- Forced alignment works best with clear, well-paced speech. Rapid speech, heavy accents, or overlapping speakers may reduce accuracy.
- Pair this node with the Add Captions processing node for automated subtitle generation with precise word timing.
- The alignment data can be consumed by downstream nodes that need timing information, such as motion graphics or video composition nodes.