ElevenLabs has built a model called Cue that tells you the exact moment each character of a transcript is spoken, and it now sits behind every speech-to-text request on the platform. The company detailed it on its blog on 28 October. It claims Cue beat every other aligner it tried across the 29 benchmark languages, and that one hour of recording takes roughly 0.5 seconds on its production GPU.
The job is called forced alignment. You already hold the text and the recording, and the aligner pins each word to its place in the sound. Select a word in one of its transcripts and playback skips to that point: an aligner is doing that. The same model times subtitles, dubbing, and lyrics for music. It has nothing to do with AI alignment in the safety sense. Scribe, ElevenLabs’ transcription model, writes the text, and Cue then sets the timing.
Length is where ElevenLabs says the open-source tools struggle. Many work well on clips of about 30 seconds, so longer files get cut into chunks, and every cut adds a little error. Over whole recordings, Cue averaged 179 milliseconds of timing error. The open-source tools landed anywhere from 283 ms to above 6,000. Cue takes a recording of any length in one pass, so there are no seams to stitch.
Messy audio narrows the gap. On English recordings with background noise, overlapping voices, or music, Cue averaged 57 ms of error, and the other tools ranged from 65 ms to 920 ms. The best rival sits only 8 ms behind. The big margins come against the weaker tools in the field.
Languages and speed are the other two claims. Open-source aligners usually ship one model per language, which leaves smaller languages thin or missing. According to ElevenLabs, just one rival, MMS-FA, covers all 29 benchmark languages, and Cue was generally quicker at finding the edges between words. On speed, the runner-up is Qwen3-ForcedAligner at nearly 2 seconds per hour of audio. The slowest tool needed well over a minute.
Music caught the team off guard. A sung word is drawn out and smeared, and few models manage it, so nobody expected Cue to do well. ElevenLabs tested it on part of the JamendoLyrics MultiLang dataset (songs sung in four languages: English, French, Spanish, German). Cue landed 92.7 percent of words inside a 0.3-second window around the reference timing, against 82.0 percent for the best alternative. Its mean start error of 0.82 seconds is about half the nearest rival’s, and it came out ahead on at least 19 of 22 songs.
All of this was scored on a benchmark ElevenLabs assembled from real requests to its own service, with the previous in-house aligner as the baseline. Danila Kozlov, the ElevenLabs researcher working on alignment, says the field has no golden benchmark, because the current aligner is already past human ability. He also says the team had to judge a character-level system with word-level labels, which “does not necessarily showcase Cue’s full potential.” No outside lab has published a comparison.
ElevenLabs also describes how fast it can fix problems. A customer reported that Georgian and a couple of other low-resource languages underperformed, and the company says it trained and shipped a better version in under 24 hours.
Developers can call Cue directly via ElevenLabs’ Forced Alignment API. Anyone who chops long recordings into chunks for an open-source aligner can send one long, noisy file through Cue and compare the timestamps at the old chunk boundaries.
Reported by the ElevenLabs blog (Mike Murphy) on 28 October 2026.
