Accuracy and language limits

Read this before you use a transcript. The page is public so you can see the limits before you log in.

Languages

This app transcribes only languages that have passed the FLEURS gate (word error at most 25%, character error at most 12%). The form currently offers cy (Welsh), en (English), fr (French), de (German), es (Spanish), pl (Polish). A language that missed the gate is not offered. The figures are in the language note shipped with this build.

Auto-detect uses the recognition engine's own language guess. The job shows that language and its probability. If the probability is below 0.5, the job is marked low confidence. A selected language does not get a probability, because the engine was told which language to use.

Welsh accuracy

Welsh Build 2 measured techiaith/whisper-large-ft-cy-en-ct2 at 19.3% word error and 8.2% character error on 40 clean-read Welsh clips from FLEURS. That figure is their development-set result. The summary file does not record a model revision, so this app does not treat a Hub pin as that measurement. This app does not repeat that run. It is not a promise about your audio.

The 10% and 5% Welsh targets are not met. Do not describe a transcript from this app as verified text, and do not claim a 5% word-error rate.

Other languages

A language stays off the form until a 40-clip FLEURS test run, scored with chapelgate-wer-norm-v1, is inside the gate above. Welsh uses the frozen Bangor measurement and was not re-scored for that decision. The 10% and 5% Welsh targets are still not met.

Transcripts keep the segments the engine returned.

What the transcript is

• Each finished transcript shows the model id, revision, and engine the recogniser reported.
• The original recognition is stored separately from your edits. Saving a correction does not overwrite it.
• SRT and VTT use the original timed text. TXT, DOCX, and PDF include the original and the correction.
• Speaker labels are not produced. Nothing in the transcript, the editor, or the exports is a diarised speaker.
• There is no translation.

Operational limits

• Upload types: mp3, wav, m4a, mp4, checked by extension, content type, and ffprobe.
• Size: 1000 to 209715200 bytes, at most 10 files at once.
• Duration: at most 14400 seconds. A file whose decoded audio would exceed 4294967296 bytes is rejected before transcription.
• Retention: jobs older than 30 days are deleted on startup and on a timer, including the audio and the transcript.
• If the recognition engine cannot be loaded, the job fails and no transcript is stored.

Review the transcript before you use it.