My first update with transcripts is now in the App Store!
It's a phased rollout that will auto-update over the next week for everyone, or you can update manually from the App Store app whenever you'd like.
More details: https://www.reddit.com/r/overcast/comments/1sg7epz/version_20264_released_transcripts/
@overcastfm Congrats!
Transcription for Portuguese works, but sentence breaks are nonexistant (likely not your fault). It appears to force a break after 60–100 words — is that your doing, or the model’s? If it’s you, a few suggestions:
• The algorithm appears to add breaks immediately after the letter sequence ends, even there’s punctuation afterwards — would be better to break at whitespace instead
• In fact, why not scan for a recent sentence break (period, exclamation or question mark followed by whitespace) rather than breaking mid-sentence? To avoid excessively short lines, it could limit to the second half of the excessively-long line (e.g. most recent 30–50 words)
• The extra effort solution: if you have timings for individual words, find the longest pause after the end of a sentence, and break there (or, if you have voice change information, break when the speaker changes)
@dluz @overcastfm I'm seeing the same thing with the transcripts of French podcasts. It must be the model being better at English and optimised for speed.
Your suggestion of reformatting for sentence breaks and line length would be a good fix to make the transcript easier to scan, but that way not be so simple to fit with the rest of the functionality.
Overcast
Daniel Luz
multigreg