Neural TTS Latency Benchmarks Across Leading APIs
Real-world TTS latency trails vendor claims by hundreds of milliseconds in production conditions.

- Written by
- The Voice Layer Editors
- Updated
- October 25, 2026
- Reading time
- 5 min read
- Sources cited
- 4 sources ↓
Time to first audio, or TTFA, is the number that decides that a voice agent feels like a conversation rather than a phone call to a broken hold line. This piece breaks down what that metric actually measures, why vendor-reported numbers diverge so sharply from third-party benchmarks, and how the leading real-time TTS APIs perform once you look past the headline figure.
- TTFA (sometimes called Time to First Byte, or TTFB) measures the elapsed time from when an API receives text to when it delivers the first audio chunk to the client. It does not measure how long the full audio takes to generate. Total generation time scales with the length of the text, and for a live agent, that number is close to irrelevant. Nobody waits for a sentence to finish rendering before they start listening to it.
- The reason TTFA carries so much weight comes down to how human conversation actually works. Turn-taking gaps between two people run somewhere between 100 and 300 milliseconds. Pushing past that window makes a pause start to read as hesitation. Pushing past roughly a second makes the interaction stop feeling slow and start feeling broken, prompting a caller to ask "hello, are you there?"
- TTFA is often confused with other metrics that measure something adjacent but different. Total generation time matters for batch jobs like audiobook rendering, where nobody's waiting on the other end. Real-Time Factor, or RTF, describes synthesis speed relative to playback speed, which tells you about headroom rather than perceived delay: an RTF of 0.3 means audio generates several times faster than it plays, which is useful for capacity planning but says nothing about how long a caller waits for the first word. Word Error Rate, or WER, measures accuracy, a completely separate axis from speed. Conflating any of these with TTFA is how a procurement conversation goes sideways before it starts.
Why the same API shows different numbers depending on measurement method
- Vendor-reported latency figures are almost always gathered under controlled conditions: dedicated test infrastructure, minimal network contention, often a single request with no concurrent load. Production traffic looks nothing like that.
- Several factors vendor numbers routinely leave out: network path and geographic distance between the client and the inference node, connection setup overhead (a WebSocket handshake behaves differently from a per-request HTTP call, and that difference can meaningfully affect latency), text normalization time before phoneme prediction even begins, and client-side buffering before playback starts.
- In a full voice agent, TTS is never the only thing standing between a user's question and a spoken answer. Speech-to-text and the LLM's first-token latency both happen first, and they compound with TTS TTFA before any sound reaches the user's ear. Benchmark data from Coval illustrates how thin the margin for error is: each component in a voice pipeline, STT, LLM first-token latency, and TTS TTFA, compounds with the others, and even a modest gap between two TTS providers can push a borderline pipeline past the point where conversation stops feeling natural.
Median latency (P50) does not decide whether users complain. P95 and P99, the tail-end percentiles, describe the calls that go wrong, and production architectures live or die on those tails, not on the average case. A useful companion metric here is the interquartile range, or IQR, which acts as a consistency proxy: a narrow IQR produces a median that fairly represents what users experience, while a wide IQR means the median flatters the provider and hides a rougher tail behind it.
How the leading real-time TTS APIs perform: provider benchmarks with sources named
- Cartesia Sonic is built around ultra-low TTFA, and its architecture is the reason why. Cartesia reports roughly 40 milliseconds TTFA for Sonic 4 and 90 milliseconds for Sonic-3 on its own site, figures made possible by a State Space Model (SSM) architecture rather than the standard transformer stack most competitors run. Third-party numbers tell a different story: the Coval benchmark records Sonic-3 at 188 milliseconds P50, with an IQR of 100 milliseconds, a spread wide enough that the median understates just how much worse the tail can get in production. That gap between Cartesia's 90ms and Coval's 188ms comes from differing measurement conditions, as described above, playing out in a single provider's numbers. The product line covers 40-plus languages and instant voice cloning from a short audio reference. Pricing runs on a credit system with tiered plans, and a per-character rate of $0.038. The obvious fit is latency-critical real-time agents, WebRTC voice bots, and anything that needs to handle mid-sentence interruption gracefully.
- ElevenLabs Flash v2.5 is positioned as the real-time, multilingual conversational model in ElevenLabs' current lineup. ElevenLabs claims roughly 75 milliseconds latency for it. Coval's benchmark puts Flash v2.5 at 288 milliseconds P50 and 304 milliseconds P75, actually slower under Coval's test conditions than ElevenLabs' own Turbo v2.5, which came in at 264 milliseconds P50 and 279 milliseconds P75 on the same test. Flash v2.5 expanded on Flash v2 by moving from English-only support to 32 languages, a meaningful jump for anyone building multilingual agents. Connection method matters here too: HTTP streaming adds that same 40-to-100-millisecond overhead mentioned earlier, whereas WebSocket multiplexing let Turbo v2.5 hit 257 milliseconds P50 under Coval's conditions, a reminder that transport choice is not a minor implementation detail but a real latency lever. Turbo v2.5 is marked deprecated on ElevenLabs' current model list as of 2026, with Flash models recommended over it across all use cases, making Flash v2.5 the active real-time choice going forward rather than Turbo. Pricing for the Flash and Turbo tier is $0.05, with subscription plans starting at $60. The strongest fit is for teams that need emotional inflection, voice cloning nuance, and dialect stability in a multilingual conversational agent.
- ElevenLabs Eleven v3 belongs in a different category entirely: expressive and high-fidelity, but not built for real-time use. It's been generally available since February 2, 2026, supporting a broad range of languages and voices, with multi-speaker dialogue and inline audio tags like [laughs], [whispering], and [sarcastic] that give it a level of expressive control the real-time models don't attempt. Coval's benchmark data on ElevenLabs' Multilingual v2 model recorded 1,232 milliseconds P50, and v3 is similarly positioned as a high-fidelity non-real-time model, a category apart from Flash and Turbo rather than a competitor to them on speed. Pricing runs at a higher per-character rate than the real-time models, with subscription API pricing also at a higher tier. The right use case here is anything rendered once and played back many times: pre-recorded IVR prompts, audiobooks, marketing voiceovers. It is not a model built for live calls, and nothing about its design suggests it's trying to be.
Methodology & sources
- https://gradium.ai/content/tts-latency-benchmark-2026
Provided the 1,232 milliseconds P50 figure for ElevenLabs Multilingual v2 used to characterize v3 as a non-real-time model.
- Best Low-Latency TTS APIs in 2026: TTFA, P99 and Pipeline Impact
Sourced the discussion of interquartile range as a consistency proxy for evaluating TTS latency distributions.
- https://www.coval.ai/blog/best-text-to-speech-providers-in-2026-how-to-choose-(and-why-vendor-benchmarks-lie)/
Provided the third-party benchmark figures for Cartesia Sonic-3, ElevenLabs Flash v2.5, and Turbo v2.5, including P50, P75, and IQR values used throughout the provider comparisons.
- https://elevenlabs.io/blog/meet-flash
Confirmed that ElevenLabs Flash v2.5 expanded language support to 32 languages compared to the English-only Flash v2.