ElevenLabs details Cue, its character-level aligner
ElevenLabs has published details of Cue, the forced-alignment model behind its Scribe transcription stack, and says its own benchmarks show Cue beating every aligner it tested on accuracy and speed.
ElevenLabs published details of Cue, a forced-alignment model that places each character of a transcript on the audio timeline. Cue is the last step in the company's speech-to-text pipeline: Scribe transcribes the audio, and Cue sets the timing. Every Scribe request goes through it. The model also handles subtitles, dubbing and music lyric timing, and developers can call it directly through the Forced Alignment API.
ElevenLabs says Cue processes recordings of any length in a single pass. Many aligners handle clips of about 30 seconds, so longer files have to be split into chunks, and each split can add timing error. On full recordings, the company reports an average timing error of 179 ms for Cue. The open-source tools it tested ranged from 283 ms to more than 6,000 ms. On noisy English recordings, Cue averaged 57 ms of error, compared with 65 ms to 920 ms for the other tools.
On speed, ElevenLabs says Cue aligns an hour of audio in about half a second on its production GPU. The next fastest tool it tested, Qwen3-ForcedAligner, took nearly 2 seconds. The company also says Cue is smaller than the aligner it replaces.
ElevenLabs built and ran all of these tests itself, on a benchmark drawn from real customer requests. Any word a tool failed to place counted as an error. The benchmark covers 29 languages, and MMS-FA was the only other tool that handled all of them. Researcher Danila Kozlov said there is no reliable ground-truth benchmark at character level, so the team scored Cue with word-level labels.
After launch, a customer reported weaker results in Georgian and some other low-resource languages. The post was cut off before it described how ElevenLabs fixed this.