| name | text-to-speech |
| description | Generate speech audio from text using multiple TTS providers — ElevenLabs, OpenAI, Google Cloud, fal.ai CSM-1B, and edge-tts (free). Smart-routes by quality, cost, and latency requirements. Use when user says "text to speech", "TTS", "generate voice", "voiceover", "narration", "read aloud", "convert text to audio", or "speech synthesis". |
Text-to-Speech
Smart-routed speech synthesis across multiple providers. See reference.md in this directory for full provider APIs, code examples, and workflows.
When to Activate
- User wants to convert text to speech audio
- Generating voiceover or narration for video/podcast
- Batch converting documents to audio
- Voice cloning or custom voice creation
- Multilingual speech synthesis
- Real-time or streaming TTS
Do NOT Use For
- Speech-to-text / transcription (use Whisper or Deepgram)
- Music or sound effects generation
Provider Routing
| Need | Provider | Cost | Latency | Quality |
|---|
| Free / prototyping / batch | edge-tts | Free | ~200ms | Good |
| Best quality / emotion | ElevenLabs | $$$ | ~75-300ms | Excellent |
| OpenAI ecosystem | OpenAI TTS | $$ | ~300ms | Very Good |
| Multilingual / enterprise | Google Cloud TTS | $ | ~100ms | Very Good |
| fal.ai ecosystem | fal.ai CSM-1B | $ | ~500ms | Good |
Decision Flow
Is this a prototype or personal use?
YES → edge-tts (free, no API key)
NO →
Need best emotional expressiveness?
YES → ElevenLabs (Eleven v3)
NO →
Already using OpenAI APIs?
YES → OpenAI TTS
NO →
Need 50+ languages or enterprise SLA?
YES → Google Cloud TTS
NO → edge-tts
Cost Comparison
| Provider | Free Tier | Paid Rate | Best Value For |
|---|
| edge-tts | Unlimited* | Free | Prototyping, personal, batch |
| ElevenLabs | 10K chars/mo | ~$0.30/1K chars | Premium content, voice cloning |
| OpenAI | None | $15-30/1M chars | OpenAI ecosystem users |
| Google Cloud | 1M chars/mo (Standard) | $4-160/1M chars | Enterprise, multilingual |
| fal.ai CSM-1B | Credits-based | ~$0.01/request | fal.ai ecosystem users |
Tips
- Start free: Use edge-tts for prototyping, switch to paid for production
- Chunk long text: Most APIs limit 1K-5K chars per request. Split at sentence boundaries
- Cache results: TTS output is deterministic — cache to avoid re-generating
- Normalize audio:
ffmpeg -af loudnorm for consistent volume
- Match voice to content: Narration → stability 0.7+, dialogue → expressiveness 0.3-0.5
For full provider setup, code examples, voice lists, and workflows, read reference.md.
