Design: Conversation Plan

Problem

Instruments generated independently with no inter-listening produce a mechanical ensemble. Real musicians listen and respond to each other.

Solution

• ConversationPlan IR: primary voice per section, conversation events (call-response, fill, tutti, solo, trade)
• Conversation Director: Step 5.5 in the pipeline, produces plan only (does NOT modify notes)
• Reactive Fills: detect phrase-end silences (gap ≥ 1.0 beat), insert idiomatic 2-4 note fills
• Frequency Clearance: symbolic collision detection (±3 semitones + time overlap), octave displacement to reduce masking
• 4 critique rules: conversation_silence, primary_voice_ambiguity, fill_absence_at_phrase_ends, frequency_collision_unresolved

ConversationEvent Types

Event	Description
call_response	One voice states, another answers
fill_in_response	Background voice fills silence
tutti	All voices play together
solo_break	Single voice featured
trade	Voices alternate short phrases

Files

• src/yao/ir/conversation.py — ConversationPlan, ConversationEvent, BarRange
• src/yao/generators/plan/conversation_director.py — generate_conversation_plan
• src/yao/generators/reactive_fills.py — detect_fill_opportunities, generate_reactive_fills
• src/yao/generators/frequency_clearance.py — symbolic collision detection
• src/yao/verify/critique/conversation_rules.py — 4 critique rules