Rate Limiting Guide for Voice AI

A common mistake when using voice AI models is misunderstanding Rate Limiting as simply 'requests per minute (RPM).' ElevenLabs' models set limits based not on request count, but on Concurrency—the number of audio generation jobs currently being processed.

Because audio generation processing time varies depending on input length and model, request count alone cannot predict actual server load. Therefore, for efficient scaling, you need to manage a Concurrency Pool and implement client-side patterns that handle HTTP 429 errors that occur when the limit is exceeded.

Key features are as follows:

• Graceful Degradation: Instead of immediately rejecting requests when the limit is reached, a Priority Queue processes requests by adding a delay of about 50ms.
• Using WebSocket: Using WebSocket means only active generation jobs count toward the limit, significantly increasing effective processing capacity.
• Per-Model Differentiated Limits: Allowed concurrency limits differ by model family—Multilingual v2, Flash, STT, etc.—so these need to be managed separately.
• Monitoring: The current-concurrent-requests and maximum-concurrent-requests headers let you check the current status in real time.