voice-audition 0.3.0

Build and maintain a rich voice catalog across TTS providers

voice-audition

The casting director for your AI voice agent. Search 670+ voices across 6 TTS providers with semantic search.

Install

pip install voice-audition

Setup

# Copy env template and fill in keys
cp .env.example .env

# Required: Moss credentials for semantic search
# Get them at https://platform.inferedge.dev
MOSS_PROJECT_ID=...
MOSS_PROJECT_KEY=...

# Optional: Provider API keys for sync + enrichment
ELEVENLABS_API_KEY=...
OPENAI_API_KEY=...
DEEPGRAM_API_KEY=...
RIME_API_KEY=...
CARTESIA_API_KEY=...
PLAYHT_API_KEY=...
PLAYHT_USER_ID=...

Quick start

# Sync voices from all providers
voice-audition sync

# Build search index
voice-audition index

# Search
voice-audition search "warm female voice for healthcare"
voice-audition search "authoritative male British accent"

# Check catalog
voice-audition stats

# Monitor provider reliability
voice-audition monitor

Commands

Command	What it does
voice-audition sync [providers...]	Sync voices from TTS providers (all if none specified)
voice-audition index	Build or rebuild the Moss semantic search index
voice-audition search <query>	Semantic search the voice catalog (--top-k for result count)
voice-audition enrich [providers...]	Enrich unenriched voices with descriptions and traits (--model to pick classifier)
voice-audition monitor	Check provider reliability via status pages
voice-audition stats	Show catalog statistics
voice-audition mcp	Start the MCP server for Claude integration

MCP Server

Add to Claude Desktop config:

{
  "mcpServers": {
    "voice-audition": {
      "command": "voice-audition",
      "args": ["mcp"],
      "env": {
        "MOSS_PROJECT_ID": "...",
        "MOSS_PROJECT_KEY": "..."
      }
    }
  }
}

Exposes 5 tools:

Tool	What it does
search_voices	Semantic search across the full catalog
get_voice	Get detailed info for a specific voice
filter_voices	Filter by gender, provider, cost, latency
get_providers	List available TTS providers and status
get_catalog_stats	Voice counts, coverage, freshness

How it works

voice-audition search "warm female for healthcare"
    |
1. Query embedded (Moss, moss-minilm model)
    |
2. Hybrid search: semantic similarity + keyword matching (alpha=0.7)
    |
3. Metadata filters applied (gender, provider, cost, latency)
    |
4. Top-k results ranked by relevance score

Voice catalog

• 670+ voices across 6 providers: ElevenLabs, OpenAI, Deepgram, Rime, Cartesia, PlayHT
• Research-backed schema: 8 perceptual traits, texture, pitch, emotional range
• Synced every 6 hours via GitHub Actions
• Provider reliability monitoring via status pages

Enrichment pipeline

Most providers ship voice metadata. Rime does not -- its 610 voices have no descriptions. The enrichment pipeline generates audio samples and classifies them with a local model (Qwen2-Audio via mlx-audio) to fill in traits and descriptions.

pip install voice-audition[enrich]
voice-audition enrich rime --model qwen2-audio

The classifier is currently stubbed and needs implementation.

Research

See the research/ directory. Key findings:

• Voice perception collapses to 2 axes: warmth vs authority (McAleer et al.)
• Speech rate is the #1 trust predictor, not pitch
• Emotional voices: +50% CSAT but -20% accuracy (Deepgram)

Development

git clone https://github.com/mnvsk97/voice-audition.git
cd voice-audition
pip install -e ".[mcp]"

License

MIT