Intel / article

ElevenLabs Cue: single-pass forced alignment across 29 languages

Why it matters

Cue aligns transcripts to audio in a single pass for any recording length, removing the that accumulates timing error. It is available through the Forced API for subtitle, dubbing and voice pipelines.

Key takeaways · AI-distilled

Most aligners handle clips of about 30 seconds, so long files get chunked and each seam can add timing error; Cue aligns whole recordings in one pass, with a 179 ms average error versus 283 ms to over 6,000 ms for the open-source tools ElevenLabs tested.
ElevenLabs says Cue aligns an hour of audio in about half a second on its production GPU, against nearly 2 seconds for Qwen3-ForcedAligner, the next fastest tool tested.
On English recordings with background noise, overlapping speech or music, Cue's average error was 57 ms, compared with 65 ms to 920 ms for the other aligners.
Cue works at the character level but was scored on word-level benchmarks built from real ElevenLabs requests, which ElevenLabs says likely understates its precision.
On a subset of JamendoLyrics songs in four languages, Cue placed 92.7% of words within 0.3 seconds of the label versus 82.0% for the next best tool, and beat every other tool on at least 19 of 22 songs.

Terms in this piece · Glossary

alignment — The work of making AI systems actually pursue what their builders and users intend, rather than something subtly or dangerously different.
chunking — Splitting documents into passages small enough to embed and retrieve individually — the step that quietly determines whether retrieval works at all.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.