Matching each character to when it was spoken
Voice AI company ElevenLabs on Oct. 28 (Oct. 29 Japan time) released "Cue," a model that aligns audio and text character by character in time. It determines the point in the audio at which each character of a transcript was spoken. This task, called "forced alignment," is essential for subtitles, dubbing and lyric display.
Cue is already built into every request as the final stage of the company's transcription processing. It is also available through a standalone "Forced Alignment API." Because it processes a recording in one pass without splitting it, it avoids the problem of timing drift at split points, the company said.
Timing error about 60% of open tools'
ElevenLabs compared it with open-source tools on recordings in 29 languages. Cue's average timing error was 179 milliseconds, compared with 283 milliseconds to more than 6,000 milliseconds for the other tools. On noisy English recordings, the average was 57 milliseconds.
Processing an hour of audio took about 0.5 seconds on the company's production GPU. Alibaba's Qwen3-ForcedAligner, from China, took about 2 seconds, and the slowest tool took more than a minute.
| Tool | Time to process 1 hour of audio (seconds) |
|---|---|
| Cue | 0.5 |
| Qwen3-ForcedAligner | 2 |
High accuracy on songs too
On songs in English, French, Spanish and German, it placed 92.7% of words within 0.3 seconds of the correct time. The next-best tool scored 82.0%.
The only compared tools named were MMS-FA, which uses Meta's multilingual speech model, and Qwen3-ForcedAligner. Pricing was not included in the announcement. All figures are from ElevenLabs' own evaluation.
