Announcing Improvements in Transcription Accuracy and Speed

In Premiere, our Speech to Text transcription model supports 18 languages. I am pleased to announce a new drop of our Speech to Text model in Premiere Beta, targeted for version 26.2. When you run transcription in Premiere, you should notice improved accuracy in languages, dialects, and overlapping dialogue.

Accuracy

Expanded language accuracy improvements: All of Premiere’s languages are showing higher accuracy. Some languages are seeing up to 36% reduction in errors.

Better handling of overlapping dialogue: You should notice more words being attributed to the correct speaker as well improvements in the change points between speakers.

English dialogue improvements: Premiere should be better at recognizing dialects for English speakers. When tested on both African American and Singaporean English data sets, Premiere has a 28% relative improvement in word error rate (tested in CORAAL and National Singapore Corpus).

Performance

Transcription wait time should be reduced, as long as you are on recommended hardware specs. The chart below shows how many seconds it takes to transcribe one hour of dialogue-heavy footage with 20 different speakers in English. Note: transcription time takes longer when Premiere separates speakers (e.g. Speaker 1, Speaker 2). We have tested on some common machines, but would like to hear if things are performant for you. 

Tell us what you think

Please test it in any languages that you support in your editing. Let us know how it’s doing with different accents, kids, and audio quality. We may want to follow up with you directly if you have sample audio for us to test.