Summary: ElevenLabs has shared details on Cue, the forced alignment model that timestamps every character in an audio file and powers its Scribe speech-to-text family. It handles recordings of any length in a single pass and aligns an hour of audio in about half a second on a production GPU. For anyone building subtitles, dubbing, karaoke, or click-to-play transcripts, the interesting part is what Cue fixes about the chunking and noise problems in older aligners.
Why forced alignment matters to your product
Forced alignment is the plumbing behind the moment a transcript highlights the word being spoken, or a subtitle lands in sync. You hand the model audio plus a known transcript, and it returns the timeline. According to ElevenLabs researcher Danila Kozlov, humans have a feel for where a word starts, but character-level, sub-millisecond precision is a different problem entirely.
Cue is the final step in every Scribe transcription request, and it also drives subtitles, dubbing, and music lyric timing. It is available standalone through the Forced Alignment API. If you ship voice products — the kind of stack covered in One API Key Now Buys Your App a Voice — this is the layer that decides whether your timestamps feel polished or broken.
The three failure modes of traditional aligners
The team interviewed customers about what frustrated them with the previous aligner, and the complaints matched limits in open-source tools:
- Length. Many aligners only work well on ~30-second clips; longer files get chunked, and every seam adds timing error. On full recordings, ElevenLabs reports Cue’s average timing error at 179 ms, versus 283 ms to over 6,000 ms for open-source tools.
- Noisy audio. On English recordings with background noise, overlapping speech, or music, Cue’s average error is 57 ms, compared with 65 ms to 920 ms for the alternatives.
- Language coverage. Most open-source aligners need one model per language. Only MMS-FA also covered all 29 benchmark languages, and Cue was generally faster at finding word boundaries.
Speed follows the same pattern: Cue aligns an hour of audio in roughly half a second, while the next fastest tool tested, Qwen3-ForcedAligner, takes nearly 2 seconds and the slowest took over a minute.
What is actually different in Cue
Cue is a single model covering more than 90 benchmark languages, and it takes recordings of any length in one pass — no chunking, no accumulated seam errors. Notably, it is smaller than the aligner it replaces, per Kozlov, while running faster on a GPU.
The team benchmarked against real-world requests from the previous aligner’s traffic, counting any unplaced word as an error. Because Cue works at character level but word-level benchmarks are the best available ground truth, those scores likely understate its precision. Kozlov noted that past human-level alignment, there is no golden benchmark left to check against.
One detail worth copying as a practice: shortly after rollout, a customer flagged that Georgian and a few other low-resource languages underperformed. ElevenLabs researched, retrained, and shipped a new model version in under 24 hours. Tight feedback loops between research and production users did that — not the model itself.
Music alignment is the sleeper feature
Sung words stretch and blur, which is why most aligners fail on music. On a subset of the Creative Commons JamendoLyrics MultiLang dataset, Cue placed 92.7% of words within 0.3 seconds of the label, versus 82.0% for the next best tool, and its average start error of 0.82 seconds was about half the closest competitor’s. It handled English, French, Spanish, and German songs in one pass.
If you want to try it
Cue already backs every ElevenLabs Speech to Text transcription, so you may be using it indirectly. To call it directly, POST to the /v1/forced-alignment endpoint with an audio file (under 1 GB per the API reference) and your transcript as plain text — diarized text is not supported. The response returns character and word timings plus a loss score you can use as an alignment confidence signal. Pricing matches the Speech to Text API, with files up to 3 GB and 10 hours of audio per the docs.
One limitation to keep in mind: the benchmark figures above come from ElevenLabs’ own comparisons, so it is reasonable to sanity-check Cue against your own audio before committing a pipeline to it.
Sources
- Forced alignment with Cue: Character-level audio timing
- Forced Alignment | ElevenLabs Documentation
- Create Forced Alignment | ElevenLabs Documentation
View the original post on X