Forced alignment

Forced alignment is a method that takes a transcript you already have and a recording of the same speech, then finds the moment each word starts and ends. The word "forced" means the text is fixed: the aligner must fit those exact words to the audio, rather than choosing its own words. That fixed text is what separates alignment from automatic speech recognition (ASR), which writes the words itself.

Alignment fits known words to sound

The aligner works from acoustic models of speech sounds. It splits the transcript into phones or words, then searches the audio for the best match between each sound and its position in time. The result is a list of start and end times, one for each unit. The Montreal Forced Aligner, an open-source tool described in a 2017 Interspeech paper, is built for this job: it aligns speech to text and times the word and phone boundaries ¹.

Recent systems combine the two steps. WhisperX adds word-level timestamps to Whisper output, using voice activity detection and forced phoneme alignment ². The recognizer writes the words, and a forced aligner then times them.

A concrete example

Suppose you have a correct transcript of a meeting, "the meeting starts at nine." Forced alignment finds when "meeting" begins in the recording and when it ends, and does the same for every other word. A captioning tool then shows each word on screen at the right moment. Because the words were fixed in advance, the aligner never has to guess them, which is why it can be more precise about timing than a recognizer that writes the words and times them at the same time.

Why it matters for dictation

Dictation puts words on the page as you speak, so the timing is already there, but it is not stored. Forced alignment gives you the timing for a text you wrote or corrected afterward. If you dictate a long memo, aligned times let you click a word and hear the matching sound, which is useful when you check a draft against the audio. It also produces the timed text that captions need.

Alignment has limits. It assumes the transcript is right, so a wrong word forces the aligner to place it somewhere, and the times around it drift. A word error rate (WER) measures how often the words are wrong, and that error is what alignment inherits. Clean up the words first, then align. For caption files, the timed output usually goes into an SRT file, and the times themselves are covered in our entry on timestamps.

Related terms

• Timestamp (transcription): the time marker that alignment produces.
• SubRip subtitle file (SRT): a caption format built from timed text.
• Speaker diarization: labeling who said each part, often done alongside alignment.

Guides that cover it

• Captions and subtitles: formats, rules and tools
• Timestamps and speaker labels in transcripts
• Transcription software: how to choose one