voice-audition

The casting director for your AI voice agent. Search 697 voices across 9 TTS providers with semantic search, run use-case auditions, and compare API vs self-hosted costs.

Install

pip install voice-audition
pip install voice-audition[enrich]   # adds Qwen2-Audio enrichment
pip install voice-audition[mcp]      # adds MCP server for Claude

Setup

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"

# Run a full audition
voice-audition audition "fertility clinic for anxious IVF patients" --gender female

# Compare costs at 100k minutes/month
voice-audition costs 100000

Commands

Cost calculator

$ voice-audition costs 100000

# Compares per-provider API pricing against self-hosted open source:
#   ElevenLabs  100k min = $3,000/mo
#   Cartesia    100k min = $1,500/mo
#   Kokoro      100k min = $0 (self-hosted, GPU cost only)
#   Piper       100k min = $0 (self-hosted, CPU-capable)

At high volume, self-hosted open source voices (Kokoro, Piper, Orpheus) can cut costs to near-zero -- the calculator shows the breakeven.

MCP Server

Add to Claude Desktop config:

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

Exposes 7 tools:

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)
   Name-based vibes fill in for unenriched voices
    |
3. Metadata filters applied (gender, provider, cost, latency)
    |
4. Top-k results ranked by relevance score

Voice catalog

697 voices across 9 providers:

• Diff-based sync with lifecycle tracking (new/deprecated/changed detection)
• Weekly pricing change detection via page hash diff
• Research-backed schema: 8 perceptual traits, texture, pitch, emotional range
• Synced every 6 hours via GitHub Actions
• Provider reliability monitoring via status pages
• Open source registry includes hosting platform data (GPU requirements, inference speed)

Enrichment pipeline

Most providers ship voice metadata. Rime does not -- its 610 voices have no descriptions. The enrichment pipeline classifies audio samples with Qwen2-Audio to fill in traits and descriptions. Tested on 10 voices.

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

Audition profiles

5 built-in use-case profiles with role-specific scoring criteria:

• Healthcare: patient comfort, trust, empathy, clarity, pacing, sensitivity
• Sales: energy, rapport, persuasiveness, confidence, resilience, likability
• Support: patience, clarity, helpfulness, professionalism, warmth, resolution focus
• Finance: authority, precision, trustworthiness, calm, professionalism, compliance
• Meditation: calm, spaciousness, grounding, non-intrusive, breath quality, presence

Self-hosting

Open source voices (Kokoro, Piper, Orpheus) run locally with no API costs. The catalog tracks GPU requirements and inference speed for each. Use voice-audition costs <minutes> to see when self-hosting beats API pricing for your volume.

Development

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

License

MIT