How ElevenLabs latency figures are measured
The latency figures on the models page are medians of the time ElevenLabs takes to respond to a request, measured on production traffic. They exclude network round-trips and the time your application spends before and after the request, so the latency your users experience is always higher.
Published latency figures for ElevenLabs models
Every figure is a median (p50). The text to speech figures are measured on WebSocket streaming requests with warm voices: a warm voice is one that has already been used for a generation.
Medians vary through the day with traffic, and are highest during peak hours. The figures on this page were last updated on October 29, 2026.
What do the text to speech latency figures measure?
The text to speech figures are server-side latency: the end-to-end time from when ElevenLabs receives a request to when it sends the first audio back. Server-side latency does not include:
- The network round-trip between your application and ElevenLabs.
- Upstream steps in your application, such as waiting for an LLM response.
- Audio player buffering before playback starts.
The latency your users experience is time-to-first-audio (TTFA), which adds these stages to server-side latency. Understanding latency describes each stage.
What does the Scribe v2 Realtime latency figure measure?
The ~150ms figure for Scribe v2 Realtime is the median time to committed transcript, taken from production metrics. Realtime transcription returns two kinds of result: partial transcripts while a segment of speech is in progress, and a committed transcript once the segment is final. See Realtime speech to text for how committing works.
Can ElevenLabs latency figures be compared with other providers?
Only when both figures measure the same thing. The ElevenLabs text to speech figures are server-side latency. They are not comparable with time-to-first-byte (TTFB) or TTFA figures measured from a client, which include the network round-trip and, for TTFA, playback buffering. To compare providers, measure TTFA for each from the same client, region and voice type.
What makes latency higher than the published figures?
The published ElevenLabs latency figures are a reference point for comparing models, not a guarantee for every request. Server-side latency is higher than the published median with:
- Longer inputs.
- Professional Voice Clones, compared with Instant Voice Clones and default voices.
- High server load, when requests queue before generation starts.
Requests are routed to the nearest serving cluster, and the round-trip to it adds to TTFA but not to the published figures.
Understanding latency explains why each of these factors matters. To measure and reduce latency in your own application, see Latency optimization.
Models without a published latency figure
ElevenLabs does not publish a latency figure for Eleven v4, Eleven v3, Eleven Multilingual v2, Scribe v2 or Scribe v2 Medical. These models are designed for quality rather than real-time use. Latency figures for the real-time models above do not apply to them.