How to improve accuracy of speaker diarization?

Other than a clean audio recording, I'd optimize the following factors:

• When using custom vocabularies: keep the list small, and provide IPA pronunciations if you can.
• When using real-time streams: two to five speakers seem best.
• When using an audio source: set the "Maximum number of speaker" to the actual number of speakers in the file.

For sources and more information:
https://aws.amazon.com/transcribe/faqs/
https://docs.aws.amazon.com/transcribe/latest/dg/diarization.html