ElevenLabs introduces Cue, a model that aligns every character with its spoken time

發生什麼事

@ElevenLabs announced Cue, a forced alignment model. Given audio and its matching transcript, it marks when each character occurs in the recording.

According to How your transcript knows where each word is spoken, Cue can process recordings of any length in one go without splitting them first. On the company's GPUs, an hour of audio takes about half a second to align. ElevenLabs built its test set from recordings by real users and says Cue was more accurate than the other tools it tested across 29 languages. Its average timing error on full recordings was 179 milliseconds; open-source tools ranged from 283 milliseconds to more than 6 seconds. These figures come from the company's own tests (🗣 self-reported).