ElevenLabs unveils Cue for character-level audio timing

ElevenLabs has introduced Cue, a forced-alignment model designed to place every character in an audio file on a precise timeline, according to a company blog post published on Oct. 28, 2026. The release matters because the model underpins transcription timing across ElevenLabs products, including subtitles, dubbing and lyric timing, while the company says it is faster and more accurate than the aligners it tested on its benchmark.

Evidence and context

In the company’s account, Cue is the final step in the speech-to-text pipeline: Scribe transcribes audio, and Cue assigns timing. ElevenLabs said every Scribe request passes through the aligner, and that the same model is also available through the Forced Alignment API. The company described forced alignment as matching supplied text to the moment it is spoken, distinct from AI alignment in the broader safety-and-values sense.

ElevenLabs said Cue is more accurate than every aligner it tested across all 29 languages in its benchmark on real-world recordings. On full recordings, the company reported average timing error of 179 milliseconds for Cue, versus 283 milliseconds to more than 6,000 milliseconds for open-source tools it compared against. On English recordings with background noise and other issues, it reported an average error of 57 milliseconds, compared with 65 milliseconds to 920 milliseconds for the other systems in the test set.

The company also said Cue handles recordings of any length in a single pass, avoiding the chunking that can add error on long files. ElevenLabs reported that the model can align an hour of audio in around half a second on its production GPU, while the next fastest tool it cited, Qwen3-ForcedAligner, took nearly 2 seconds and the slowest tool tested took more than a minute.

Unlike most competitors, ElevenLabs said Cue is a single model covering more than 90 benchmark languages. It added that the model works at character level rather than the word level used in many benchmarks, which the company said means the reported scores likely understate its precision. The post also said the team built Cue around real-world customer requests and that it can be updated quickly in response to user feedback; ElevenLabs cited one case in which a language-quality issue was identified, improved and re-released in less than 24 hours.

The company said Cue also performed well on a subset of songs from the JamendoLyrics MultiLang dataset, aligning English, French, Spanish and German lyrics in a single pass. ElevenLabs reported that 92.7% of words were placed within 0.3 seconds of the label, compared with 82.0% for the next best tool, and that Cue was more accurate than the other tools on at least 19 of the 22 songs in the test subset.

ElevenLabs had previously added a forced-alignment endpoint on April 7, 2025, according to its changelog. Separate company material published on Oct. 5, 2026, explained the difference between native word-level transcription and a separate forced-alignment pass. The new Cue post builds on that product line, but the company’s benchmark figures are self-reported and limited to its described test conditions.