ElevenLabs Cue aligner places every character on the timeline, aligning one hour of audio in about half a second

ElevenLabs has detailed Cue, a forced alignment model that places every character of a transcript on an audio timeline. According to the company, Cue is more accurate than every aligner it tested on real-world recordings in all 29 languages of its benchmark, and it aligns one hour of audio in around half a second on its production GPU.

Where Cue sits in the stack

Cue is the last step of the speech-to-text pipeline. Scribe, ElevenLabs' Speech to Text model, transcribes the audio, and Cue sets the timing. Every Scribe request passes through it. ElevenLabs says the same model also powers subtitles, dubbing and lyric timing for music, and is available on its own through the Forced Alignment API. Pricing is not yet disclosed.

Forced alignment here means matching known text to the moment each word is spoken. It is unrelated to AI alignment in the sense of matching model behavior to human values.

Benchmark results (ElevenLabs' own tests)

ElevenLabs built the benchmark from real-world requests, using its previous aligner as a baseline, and scored every tool the same way. Any word a tool failed to place counted as an error. The figures below are the company's claims.

• Full-length recordings: Cue's average timing error is 179 ms. The open-source tools tested ranged from 283 ms to more than 6,000 ms.
• Noisy English recordings (background noise, overlapping speech, ambient music): Cue averages 57 ms error, against 65 ms to 920 ms for the others.
• Speed: Cue aligns one hour of audio in around half a second. The next fastest tool, Qwen3-ForcedAligner, takes nearly 2 seconds. The slowest tested took well over a minute.
• Languages: MMS-FA is the only other tool that handles all 29 benchmark languages. ElevenLabs says Cue was generally quicker at finding word boundaries than MMS-FA. The company describes Cue as a single model covering more than 90 languages, whereas most open-source aligners need a separate model per language.

Design choices

Many aligners work well on clips of about 30 seconds, so longer files must be split into chunks. ElevenLabs says each chunk can add timing error, and stitching the seams adds more. Cue takes recordings of any length in a single pass. The company says Cue is smaller than its previous aligner and still runs on a GPU.

Cue works at the character level, but most benchmarks use word-level labels. ElevenLabs researcher Danila Kozlov said there is no golden benchmark for this, and that word-level benchmarks "do not necessarily showcase Cue's full potential." The company says its scores likely understate Cue's precision.

Singing

ElevenLabs tested Cue on a subset of songs from the Creative Commons-licensed JamendoLyrics MultiLang dataset, in English, French, Spanish and German, aligned in one pass. Cue placed 92.7% of words within 0.3 seconds of the label, compared with 82.0% for the next best tool. Its average start error was 0.82 seconds, about half that of the closest competitor. It was more accurate than every other tool on at least 19 of the 22 songs.

Fix turnaround

After Cue's rollout, a customer reported that Georgian and a couple of other lower-resource languages performed worse than expected. ElevenLabs says it researched improvements, trained a new version and released it within 24 hours of the report.

What this means

For developers building on word-level timestamps, such as subtitle tools, clickable transcripts or dubbing pipelines, the main claims are about long files and noisy audio. These are the cases where ElevenLabs says chunking-based aligners accumulate error. Single-pass processing of any length and about half a second per hour of audio are the figures to test against existing workloads. All comparisons come from a benchmark ElevenLabs built and ran itself, using word-level labels for a character-level system, so independent evaluation on your own audio is the way to confirm them. The song-alignment results cover only a subset of 22 songs in four languages.