search

rishikanthc / Scriberr

Self-hosted AI audio transcription
https://scriberr.app
MIT License
457 stars 17 forks source link

Losing pyannote/whisper on template restart in Unraid #34

Closed shadowsoze closed 20 hours ago

shadowsoze commented 5 days ago

Howdy, so i can install everything easily using the updated template pointed to the nightly release, however if i restart the template and try and use it again, i'm getting this error:

DEV MODE -----> false
stderr: Traceback (most recent call last):
  File "/app/./diarize/local.py", line 2, in <module>
    from pyannote.audio import Pipeline
ModuleNotFoundError: No module named 'pyannote'

Not sure exactly why it's being deleted, it's almost as if nothing is being saved to the folders i define it to while setting up, because i can't find the whisper files either.

image

Any ideas?

rishikanthc commented 4 days ago

The nightly is behind. Use 0.3.0 that's the currently tested working one. Either ways what you're experiencing is weird. Could you share the docker compose you used please ? When the app started did the configuration wizard run ?

shadowsoze commented 1 day ago

So looks like pointing to 0.3.0 instead of nightly fixed the issue...i've done a few restarts/stopping/starting of the docker template and functionality is retained with no issue, so i think i'm ok on that one. I also had to add PUID/PGID/UMASK info to the template but that's moreso on the template creator then the actual app so i'll let them know.

The only issue i have now is diarization is hit or miss...but i think that's moreso of my source material (i've been testing with random youtube clips but they have background music so i'm assuming that's causing issues, i've done a few other voice files with just voice and it works well).

Only other question i have is using LocalAI (https://localai.io/) and point to it using it's API instead of a seperate ollama process, in whether or not it's supported? I tried pointing to a llama model but it wasn't working for me, but i think the error was related to my hardware.

rishikanthc commented 1 day ago

Ollama is supported. Anything that uses a similar api will work. The diarization can be improved and I'm working on adding a couple different algorithms to improve things. But yes stuff like background noise will throw it off