ElevenLabs has released Cue, a forced-alignment model that matches a written transcript to a recording so that each word — and, with Cue, each character — is placed at the exact moment it is spoken. In practice, that means a listener can click a word in a transcript and jump straight to that point in the audio, rather than scrubbing through a recording to find it. The model is smaller than ElevenLabs’ previous aligner, still runs on a graphics processor, and handles recordings of any length in a single pass, avoiding the timing errors that can creep in when long audio is split into chunks.
What forced alignment does
Forced alignment is a timing task, not a transcription task. A speech-to-text system first turns audio into words; alignment then works out when each of those words was spoken. That timing layer is what makes transcripts clickable, subtitles appear on cue, and dubbed dialogue land in sync with the speaker on screen.
Cue works at character level, which is finer than the word-level labels used in most benchmarks. ElevenLabs says that gap means benchmark scores probably understate how precise the model is. The model is also the final step in the company’s Scribe speech-to-text pipeline, running after transcription itself, and it already sits behind every Speech to Text request on the platform.
Benchmark results, as reported by ElevenLabs
ElevenLabs reports that Cue was more accurate than every aligner it tested, in all 29 languages. A single Cue model covers more than 90 languages, the company says, while most open-source aligners need a model for each language. On recordings, its average timing error was 179 milliseconds, versus 283 milliseconds to more than 6,000 milliseconds for the open-source tools compared.
On English recordings with overlapping speech and music, Cue’s error was 57 milliseconds, versus 65 to 920 milliseconds for rivals. It aligned one hour of audio in under a second — half a second on ElevenLabs’ graphics processor — against nearly two seconds for the next model, Qwen3-ForcedAligner, and well over a minute for the slowest.
Cue was tested on songs, using the JamendoLyrics MultiLang subset in English, French, Spanish and German. It placed 92.7% of words within 0.3 seconds of the label, against 82.0% for the next-best system, with an average start error of 0.82 seconds — about half the closest competitor’s — and it was more accurate on at least 19 of 22 songs.
Where Cue shows up
Beyond Scribe, ElevenLabs says Cue powers subtitles, dubbing and lyric timing for music. Developers can use the model through a Forced Alignment API, although the company has not stated pricing for access. The timing work sits alongside ElevenLabs’ audio tools for video, including its instant soundtracks, but Cue itself does not generate music — it places words against existing sound.
A fast fix for weaker languages
The rollout was not entirely smooth. After release, a customer reported weaker performance for Georgian and a couple of lower-resource languages. ElevenLabs says it researched the problem, trained a new version and released it in under 24 hours — a turnaround the company presents as evidence that a single, smaller model can be updated quickly when a weakness surfaces.
Why It Matters
Character-level timing sounds like a footnote, but it changes how audio can be used. Clickable transcripts turn recordings into material readers can search and quote, while tighter subtitles and dubbing make video and music accessible across languages. For developers, alignment that fast removes a bottleneck from transcription, captioning and dubbing workflows.
The benchmark claims still deserve the usual caution: the figures are ElevenLabs’ own, measured against its chosen comparisons, and real-world accuracy will vary by language, recording quality and speaker overlap. The Georgian fix, though, is a useful signal — a model that can be corrected in a day is easier to trust than one whose weaknesses linger between releases.
Background
Forced alignment has long been a quiet step inside transcription tools, usually handled by separate, language-specific models. Cue folds that step into a single model covering more than 90 languages, and ElevenLabs has made it the default timing layer behind its speech-to-text service.
Source: ElevenLabs blog
0 Comments