Force-aligned
Forced alignment (audio-text alignment)
The process of aligning a known transcript to audio to produce precise word-level timestamps. Used to upgrade sentence-level timing to word-level.
In depth
Forced alignment (also called audio-text alignment) takes a known transcript and an audio file, then computes the exact timestamp where each word or phoneme occurs. Unlike ASR (which guesses the words), forced alignment already knows the text — it only needs to find where each word starts and ends in the audio. This produces much more precise word-level timestamps than ASR alone, typically ±20ms accuracy on clean speech.
When to use it
Use forced alignment when you have a correct transcript but need word-level timestamps — for karaoke captions, word-by-word reveal animations, or upgrading sentence-level SRT to word-level timing. WhisperX uses forced alignment (via wav2vec2) to produce word-level timestamps from Whisper's segment output.
Frequently asked
How is forced alignment different from transcription?+
Transcription (ASR) guesses what words were said. Forced alignment already knows the words — it only computes when each word occurs in the audio. Alignment is more precise because it's solving a simpler problem.
What tools do forced alignment?+
WhisperX (uses wav2vec2 for alignment), Montreal Forced Aligner (MFA), Gentle, and Aeneas. WhisperX is the most common in modern caption pipelines.
Can I force-align in any language?+
Yes, if a phoneme model exists for that language. wav2vec2 covers 50+ languages. MFA supports even more with community-contributed models.
Timing data that marks the start and end of every word, not just every cue. The foundation for karaoke captions and word-by-word reveal animations.
OpenAI's open-source automatic speech recognition model. The de facto baseline for AI subtitle generation and the engine behind most modern caption tools.
The process of identifying which speaker said which words in an audio recording. Critical for interviews, podcasts, and any multi-speaker content.