[SWML Update] - asr_speaker_affinity, asr_diarize, llm_diarize_aware

[hey-august]

Type Of Change

• New SWML Method
• Update to existing SWML Method

Are These Changes Already In Production?

• Yes
• No

Description

[Provide a detailed description of the changes]

• Method(s) Added/Updated:
  • Description of what was Added/Updated on the Method(s)
  • If this is a New Method,
    • Provide a description of its purpose and functionality.
    • Provide what valid types it takes.
• Method Field(s) Added/Updated:
  • Description of what Field was changed
  • If this is a New Field
    • Provide a description of its purpose and functionality.
    • Provide what valid types it takes.

Context and Motivation

[Explain the context and the motivation behind these changes]

SWML Example (Optional)

{
  "version": "1.0.0",
  "sections": {
    "main": []
    }
}

[anthmFS]

asr_diarize (passive) will break up the transcripts to have a prefix of speaker 1, speaker 2 etc.

asr_speaker_affinity (activates asr_diarize) Whoever talks to it first will be responded to and other people will be ignored. (subject to it actually working well depends on voices etc)

llm_diarize_aware (activates asr_diarize) The LLM will be involved in the diarization, so you can tell it your name and things like that and it will remember who you are later based on what speaker said what. Hidden prompting is injected into the master system prompt and the speaker labels are fed back to the LLM.