Conversation

Notices

1. Embed this notice
   picofarad (picofarad@noauthority.social)'s status on Monday, 27-May-2024 12:02:29 JST picofarad

   https://github.com/transcriptionstream/transcriptionstream

   hmmm, will have to see

   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Monday, 27-May-2024 12:02:28 JST iced depresso

     in reply to
     @picofarad ... hmm.
     
     there are a lot of errors in the whisper/vosk transcriptions i take of myself.
   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Monday, 27-May-2024 12:05:40 JST iced depresso

     in reply to
     @picofarad i trust mistral to summarize editorial content. i don't know about this prompt, but "summarize the following text" has worked well for me
   • Embed this notice
     picofarad (picofarad@noauthority.social)'s status on Monday, 27-May-2024 12:05:41 JST picofarad

     in reply to

     • iced depresso

     @icedquinn yes, i hope someday they release better models but for now it's have some errors and a searchable corpus or zero errors and no searchable corpus.

     diarization hasn't worked for me, yet. I'm willing to try anything once, although using docker but also requiring .sh runnery is almost a non-starter.

   • Embed this notice
     iced depresso (icedquinn@blob.cat)'s status on Monday, 27-May-2024 12:09:46 JST iced depresso

     in reply to
     @picofarad just chunking input and doing summary of summaries is a normal thing.
     
     there are some tricks people have done like nltk/scapy to split sentences to make sure the chunks don't just cut in the middle of a word
   • Embed this notice
     picofarad (picofarad@noauthority.social)'s status on Monday, 27-May-2024 12:09:47 JST picofarad

     in reply to

     • iced depresso

     @icedquinn there was some commentary that mixtral (with an x) may be better since it *may* have the full 32k context window and not require moving window batching.

     I've never tried summarizing, but i wonder how hard it would be with this one corpus i have, maybe split it on a phone number or something and summarize each section.