voice-blend

voice-blend

Combine multiple voice profiles with weighted mixing to create hybrid voices.

Triggers

Alternate expressions and non-obvious activations (primary phrases are matched automatically from the skill description):

• "70/30 [voice A]/[voice B]" -> ratio-based voice blending
• "blend voices" -> voice mixing

Behavior

When triggered, this skill:

1. Loads source voice profiles from:

   • Built-in templates (voices/templates/)
   • Project voices (.aiwg/voices/)
   • User voices (~/.config/aiwg/voices/)

2. Parses blend specification:

   • Equal blend: "blend X and Y" -> 50%/50%
   • Weighted blend: "70% X, 30% Y" -> explicit weights
   • Multi-voice: "blend X, Y, and Z" -> equal thirds

3. Interpolates dimensions:

   • Weighted average of tone values (formality, confidence, etc.)
   • Merged vocabulary lists (union of prefer, intersection of avoid)
   • Dominant structure patterns from highest-weighted voice

4. Generates hybrid profile with clear lineage tracking

Usage Examples

Equal Blend

User: "Blend technical-authority and friendly-explainer"

Result: 50/50 blend
- formality: 0.45 (avg of 0.7 and 0.2)
- confidence: 0.8 (avg of 0.9 and 0.7)
- warmth: 0.55 (avg of 0.3 and 0.8)
- vocabulary: merged from both

Weighted Blend

User: "80% executive-brief, 20% casual-conversational"

Result: Weighted blend
- formality: 0.7 (0.8*0.85 + 0.2*0.15)
- confidence: 0.86 (0.8*0.9 + 0.2*0.6)
- Dominant structure from executive-brief

Multi-Voice Blend

User: "Combine technical-authority, friendly-explainer, and executive-brief"

Result: Equal thirds (33.3% each)
- All dimensions averaged across three profiles
- Vocabulary merged from all three

Blend Algorithm

Dimension Interpolation

For each tone dimension:

blended_value = sum(weight_i * value_i) / sum(weight_i)

Vocabulary Merging

• prefer: Union of all prefer lists, deduplicated
• avoid: Intersection of all avoid lists (only avoid if ALL sources avoid)
• signature_phrases: Top N from each source (weighted by blend ratio)

Structure Resolution

Structure settings use the dominant voice (highest weight):

• sentence_length: from dominant
• paragraph_length: from dominant
• use_lists: from dominant
• use_examples: averaged (rarely=0, when-appropriate=1, frequently=2)

Perspective Handling

• person: Majority vote (tie goes to second person)
• voice: Active unless all sources use passive
• tense: Present unless all sources use past

Output Format

name: technical-friendly-blend
version: 1.0.0
description: Blended voice profile
blend_sources:
  - name: technical-authority
    weight: 0.7
  - name: friendly-explainer
    weight: 0.3
tone:
  formality: 0.55  # interpolated
  confidence: 0.84 # interpolated
  warmth: 0.45     # interpolated
  energy: 0.49     # interpolated
  complexity: 0.65 # interpolated
vocabulary:
  prefer:
    - precise technical terminology  # from technical
    - concrete examples              # from friendly
  avoid:
    - marketing superlatives         # common to both
  signature_phrases:
    - "The system handles..."        # from technical (70%)
    - "Think of it like..."          # from friendly (30%)
structure:
  sentence_length: medium            # from dominant (technical)
  use_examples: frequently           # averaged

Output Location

Blended profiles are saved to:

1. .aiwg/voices/{name}.yaml (default)
2. Custom path with --output flag
3. ~/.config/aiwg/voices/ with --global flag

Integration

• Input: Takes profiles created by voice-create or built-in templates
• Output: Creates profiles usable by voice-apply
• Chain: voice-analyze -> voice-blend -> voice-apply

CLI Usage

# Equal blend of two voices
python voice_blender.py --voices "technical-authority,friendly-explainer"

# Weighted blend
python voice_blender.py --voices "technical-authority:0.7,friendly-explainer:0.3"

# Custom output name
python voice_blender.py --voices "..." --name my-hybrid-voice

# Output to specific directory
python voice_blender.py --voices "..." --output .aiwg/voices/

# JSON output for inspection
python voice_blender.py --voices "..." --json

AI Tells Handling in Blends

When blending profiles, the AI tells avoidance rules are treated as hard constraints, not blendable dimensions:

• vocabulary.avoid: The union of all source profiles' avoid lists is used, plus the full banned AI vocabulary list from ../voice-apply/references/ai-tells.md is always included. AI tells are never averaged away.
• anti_ai_tells: If any source profile includes anti-AI-tells rules, the strictest setting wins. If no source profile has them, they are added automatically from the tells reference.
• punctuation: Em dash avoidance is always enforced in blended output regardless of source profile settings.
• rhythm: Sentence length variance and paragraph length variance are always set to "high" in blended profiles. Uniform rhythms are an AI tell that blending can accidentally amplify.

The blended profile always includes the full anti_ai_tells block from voice-create's output format.

Error Handling

• Profile not found: Lists available profiles and suggests closest match
• Invalid weights: Normalizes weights to sum to 1.0
• Incompatible profiles: Warns but proceeds with best-effort blend

References

• AI Writing Tells: ../voice-apply/references/ai-tells.md