Test Out New AssemblyAI Model (Universal 3.5-Pro)

[Mephistic]

Summary

The new AssemblyAI model supports up to 30 speakers with speaker diarization (up from our original ~10).

This is a substantial lift in terms of our ability to potentially map voice lines with speakers in our hearing transcripts - we should test out the new model to determine:

• Does the speaker diarization now work properly (First level: Does it distinguish the speakers correctly? Second level: Can it name the legislators (whose names are known via committee/agenda)?)
  • We can provide a list of known names to help improve accuracy - we should send in the list of all legislators on the Committee holding the hearing at minimum (possibly there's more specific data in the agenda)
• Cost Difference (if any)
• Compare to the transcriptions/<TRANSCRIPT_ID>/utterances to see how different the diarization actually is between versions?

We would have to re-process all of our hearings with the new model to get this if it proves useful - so we should determine how much of a lift it really is in terms of data quality.