ElevenLabs unveils Cue, a tool that matches every word to its exact moment

ElevenLabs has introduced Cue, a tool that matches every character of a transcript to the exact moment it is spoken. The company says Cue can line up an hour of audio with its text in about half a second. It also says Cue beat every other aligner it tested on real-world recordings, across 29 languages.

What it does

The job is called forced alignment. Someone hands you a video and a transcript, and you match each word to the moment it is said. When you click a word in an ElevenLabs transcript and the player jumps to that spot, an aligner is doing the work. Cue is the last step after the company's Scribe model writes the text. It also powers subtitles, dubbing and lyric timing for music.

The numbers

These results come from ElevenLabs' own tests, using a benchmark built from real-world requests. Here is what the company reports:

• Full recordings: Cue's average timing error was 179 ms. Open-source tools ranged from 283 ms to over 6,000 ms.
• Noisy English recordings: Cue averaged 57 ms of error, against 65 ms to 920 ms for others.
• Speed: about half a second per hour of audio. The next fastest tool, Qwen3-ForcedAligner, took nearly 2 seconds.

A millisecond is one thousandth of a second. So the gaps are small, but they add up when a subtitle drifts over a long video.

Why length and noise matter

Many aligners work well on clips of about 30 seconds. Longer files get chopped into chunks, and each cut can add a little timing error. Cue takes recordings of any length in one pass, which avoids that. Background noise, overlapping speech and music also make many tools lose their place. Cue was built around the messy audio people actually upload.

Languages and size

Most open-source aligners need a separate model for each language, so languages with less training data often do worse or are not supported. Cue is a single model. MMS-FA is the only other tool in the test that handles all 29 languages, and Cue was generally quicker at finding the boundaries between words. Cue is also smaller than ElevenLabs' previous aligner, though it still runs on a GPU, a type of chip used for heavy computing.

What is still open

There is no independent check yet. Researcher Danila Kozlov noted there is no gold-standard benchmark for this work. The team scored a character-level system with word-level tests, which he said may not show Cue's full potential. After launch, a customer reported that Georgian and a couple of other languages with fewer resources performed worse than expected. Cue is also available on its own through the Forced Alignment API.