When building conversational agents or real-time voice applications, latency is the defining metric. If Time-to-First-Byte (TTFB) exceeds 200ms, natural turn-taking breaks down and conversational interruption becomes clunky.
We recently recorded and aggregated median latency and pricing metrics across the primary streaming Text-to-Speech APIs using WebSocket connections (US-East endpoints, median across 1,000 requests).
Summary Table
| Provider | Model | TTFB Latency | Pricing / 1M Chars | Real-Time Suitability |
|---|---|---|---|---|
| Cartesia | Sonic-3 | 85ms | $20.00 | Excellent (Fastest turn-taking) |
| Deepgram | Aura-2 | 115ms | $15.00 | Very Good (Lowest bulk cost) |
| ElevenLabs | Flash v2.5 | 135ms | $25.00 | Best Voice Realism |
| PlayHT | PlayDialog | 180ms | $25.00 | Good |
| OpenAI | TTS-1 | 240ms | $15.00 | Slower (Chunked HTTP) |
Key Takeaways for Developers
- Ultra-low latency: Cartesia Sonic-3 remains the fastest streaming engine for handling real-time interruptions and WebRTC voice bots.
- Cost at scale: Deepgram Aura-2 provides the most competitive price per character when deploying high-volume voice automation pipelines.
- Voice acting & naturalness: ElevenLabs Flash v2.5 remains the benchmark for emotional inflection, voice cloning nuance, and dialect stability.
The full benchmark dataset, methodology, and continuous updates are documented at VoiceAIBench.
Raw data and contributions are also maintained on GitHub: awesome-voice-ai-latency.
Top comments (0)