10/11/2026
10/11/2026
10/11/2026
Neuphonic TTS latency: how we measure time to first audio (p50/p90)
Neuphonic TTS latency: how we measure time to first audio (p50/p90)
Neuphonic TTS latency: how we measure time to first audio (p50/p90)
Time to first audible speech, including network time and leading silence. Neuphonic’s test from London on 12 October 2026, with 432 requests per provider.
Time to first audible speech, including network time and leading silence. Neuphonic’s test from London on 12 October 2026, with 432 requests per provider.
Time to first audible speech, including network time and leading silence. Neuphonic’s test from London on 12 October 2026, with 432 requests per provider.
Our models are built to be smaller and more efficient: being able to run on low powered CPU hardware whilst also able to benefit from parallelisable efficiency gains from GPUs.
To highlight this, on the 12th of October 2026, we measured latency from a client in London to our European deployment (hosting our frontier model NeuTTS-Max), reaching first audible speech in a median (p50) of 83 ms and a p90 of 95 ms across 432 requests, counting any leading silence.
Our models are built to be smaller and more efficient: being able to run on low powered CPU hardware whilst also able to benefit from parallelisable efficiency gains from GPUs.
To highlight this, on the 12th of October 2026, we measured latency from a client in London to our European deployment (hosting our frontier model NeuTTS-Max), reaching first audible speech in a median (p50) of 83 ms and a p90 of 95 ms across 432 requests, counting any leading silence.
Date (UTC) | Provider | Model | p50 | p90 | p95 |
|---|---|---|---|---|---|
| 12 Oct 2026 | Neuphonic | NeuTTS-Max | 83 ms | 95 ms | 99 ms |
| 12 Oct 2026 | ElevenLabs | Turbo v2.5 | 143 ms | 267.9 ms | 284.9 ms |
| 12 Oct 2026 | Cartesia | Sonic 3.5 | 231.5 ms | 284.8 ms | 316.8 ms |
Date (UTC) | Provider | Model | p50 | p90 | p95 |
|---|---|---|---|---|---|
| 12 Oct 2026 | Neuphonic | NeuTTS-Max | 83 ms | 95 ms | 99 ms |
| 12 Oct 2026 | ElevenLabs | Turbo v2.5 | 143 ms | 267.9 ms | 284.9 ms |
| 12 Oct 2026 | Cartesia | Sonic 3.5 | 231.5 ms | 284.8 ms | 316.8 ms |
Time to speech = time to first byte + leading silence. Percentiles use linear interpolation between ranked results.
What does each latency number count?
What does each latency number count?
Time to speech is the delay a listener hears: from sending text to the first audible sound. It is the number we report. Two other metrics are common, and they can differ by a lot.
Time to speech is the delay a listener hears: from sending text to the first audible sound. It is the number we report. Two other metrics are common, and they can differ by a lot.
Metric | Clock starts | Clock stops | Includes leading silence? |
|---|---|---|---|
| Time to first byte (TTFB) | Request sent | First byte of any response arrives | No |
| Time to first audio byte (TTFAB) | Request sent | First byte of audio arrives | No |
| Time to speech | Request sent | First audible audio arrives | Yes |
Metric | Clock starts | Clock stops | Includes leading silence? |
|---|---|---|---|
| Time to first byte (TTFB) | Request sent | First byte of any response arrives | No |
| Time to first audio byte (TTFAB) | Request sent | First byte of audio arrives | No |
| Time to speech | Request sent | First audible audio arrives | Yes |
In short: time to speech = time to first byte + leading silence. All three include network time from your client to the server.
In short: time to speech = time to first byte + leading silence. All three include network time from your client to the server.
Are these numbers guaranteed?
Are these numbers guaranteed?
No. They’re measured percentiles from one client location on one day, not a service-level guarantee. Your latency depends on where your code runs, what GPU the models are running on, and how you call the API. We only publish percentiles we have measured in a setting matching the deployments our customers experience.
No. They’re measured percentiles from one client location on one day, not a service-level guarantee. Your latency depends on where your code runs, what GPU the models are running on, and how you call the API. We only publish percentiles we have measured in a setting matching the deployments our customers experience.
Why does Neuphonic count leading silence?
Why does Neuphonic count leading silence?
Because the listener waits through it. A provider can send its first audio byte quickly while that audio is still silent. A benchmark that stops the clock at the first byte would miss that wait. NeuTTS-Max had a median leading silence of 0 ms, so for NeuTTS-Max its first audio byte and time to speech are close to the same. Counting silence keeps the comparison fair across providers that behave differently.
Because the listener waits through it. A provider can send its first audio byte quickly while that audio is still silent. A benchmark that stops the clock at the first byte would miss that wait. NeuTTS-Max had a median leading silence of 0 ms, so for NeuTTS-Max its first audio byte and time to speech are close to the same. Counting silence keeps the comparison fair across providers that behave differently.
What adds TTS latency on my side?
What adds TTS latency on my side?
What adds TTS latency on my side?
Most of the gap between what users see comes from a number of things you can control, and a couple things you unfortunately can’t! The following are a number of key issues:
• Region: network distance adds to every request. Measure from where your agent runs, not from your laptop.
• Serving Framework: software on the model-hosting side (i.e. in the data centre) can be optimised for faster inference. This impacts how quickly the model generates the response.
• Length of Generation: before generating any audio, the model runs a prefill step that processes the entire input text at once. Prefill cost grows with input length, and faster than linearly because attention compares every token with every other token. So a 100-word request has a longer time to first audio than a 3-word one, before generation even begins.
Most of the gap between what users see comes from a number of things you can control, and a couple things you unfortunately can’t! The following are a number of key issues:
• Region: network distance adds to every request. Measure from where your agent runs, not from your laptop.
• Serving Framework: software on the model-hosting side (i.e. in the data centre) can be optimised for faster inference. This impacts how quickly the model generates the response.
• Length of Generation: before generating any audio, the model runs a prefill step that processes the entire input text at once. Prefill cost grows with input length, and faster than linearly because attention compares every token with every other token. So a 100-word request has a longer time to first audio than a 3-word one, before generation even begins.
How does Neuphonic measure TTS latency?
How does Neuphonic measure TTS latency?
We send short English sentences to each provider on a fixed schedule from one client, time each request to the first audible speech, and report percentiles of the successful requests. The 12 October 2026 round used this set-up:
We send short English sentences to each provider on a fixed schedule from one client, time each request to the first audible speech, and report percentiles of the successful requests. The 12 October 2026 round used this set-up:
Setting | 12 Oct 2026 round |
|---|---|
| Client location | London, United Kingdom |
| Server | Neuphonic’s European deployment; each competitor’s public API |
| Cadence | 3 requests per provider every 10 minutes, 00:00 to 23:50 UTC (144 rounds) |
| Sample size | 432 requests per provider, 1,296 in total |
| Statistic | p50, p90 and p95 of successful requests |
| Text set | 10 English sentences, 44 to 84 characters (65 on average) |
| Leading silence | Measured in 10 ms steps from the start of the audio |
Setting | 12 Oct 2026 round |
|---|---|
| Client location | London, United Kingdom |
| Server | Neuphonic’s European deployment; each competitor’s public API |
| Cadence | 3 requests per provider every 10 minutes, 00:00 to 23:50 UTC (144 rounds) |
| Sample size | 432 requests per provider, 1,296 in total |
| Statistic | p50, p90 and p95 of successful requests |
| Text set | 10 English sentences, 44 to 84 characters (65 on average) |
| Leading silence | Measured in 10 ms steps from the start of the audio |
#lowlatency
#lowlatency
#lowlatency
#texttospeech
#texttospeech
#texttospeech
#voiceAI
#voiceAI
#voiceAI
#AIbenchmarks
#AIbenchmarks
#AIbenchmarks
FAQ
Everything you need to know about how we measure latency.
In Neuphonic's test on 12 October 2026, NeuTTS-Max reached first audible speech in a median (p50) of 83 ms across 432 requests, measured from London to Neuphonic's European deployment. The figure includes leading silence.
The p90 was 95 ms and the p95 99 ms, in the same 12 October 2026 test. That was the lowest p90 of the three models shown.
From a client in London, United Kingdom, sending requests to Neuphonic's European deployment. Results from other locations will be added as new rows with their own dates.
Yes, both. Time to speech is counted from sending the request to the first audible audio, so it includes the network round trip and any silence before the first word.
Both stream audio as it is generated. The docs describe SSE as giving the lowest latency for a single request; for an agent that speaks many turns, keep one websocket open per call and send <RESET> before each new turn.
Yes. Our open benchmark script measures time to first audio byte (TTFAB) across providers from wherever you run it. Run it from the region where your agent will run to see the latency your users will get.
FAQ
Everything you need to know about how we measure latency.
In Neuphonic's test on 12 October 2026, NeuTTS-Max reached first audible speech in a median (p50) of 83 ms across 432 requests, measured from London to Neuphonic's European deployment. The figure includes leading silence.
The p90 was 95 ms and the p95 99 ms, in the same 12 October 2026 test. That was the lowest p90 of the three models shown.
From a client in London, United Kingdom, sending requests to Neuphonic's European deployment. Results from other locations will be added as new rows with their own dates.
Yes, both. Time to speech is counted from sending the request to the first audible audio, so it includes the network round trip and any silence before the first word.
Both stream audio as it is generated. The docs describe SSE as giving the lowest latency for a single request; for an agent that speaks many turns, keep one websocket open per call and send <RESET> before each new turn.
Yes. Our open benchmark script measures time to first audio byte (TTFAB) across providers from wherever you run it. Run it from the region where your agent will run to see the latency your users will get.
FAQ
Everything you need to know about how we measure latency.
In Neuphonic's test on 12 October 2026, NeuTTS-Max reached first audible speech in a median (p50) of 83 ms across 432 requests, measured from London to Neuphonic's European deployment. The figure includes leading silence.
The p90 was 95 ms and the p95 99 ms, in the same 12 October 2026 test. That was the lowest p90 of the three models shown.
From a client in London, United Kingdom, sending requests to Neuphonic's European deployment. Results from other locations will be added as new rows with their own dates.
Yes, both. Time to speech is counted from sending the request to the first audible audio, so it includes the network round trip and any silence before the first word.
Both stream audio as it is generated. The docs describe SSE as giving the lowest latency for a single request; for an agent that speaks many turns, keep one websocket open per call and send <RESET> before each new turn.
Yes. Our open benchmark script measures time to first audio byte (TTFAB) across providers from wherever you run it. Run it from the region where your agent will run to see the latency your users will get.
Bring Neuphonic into your product.
Building something with voice AI? Talk to us about your use case,
deployment requirements, and how Neuphonic can fit into your stack.
Building something with voice AI? Talk to us about your use case,
deployment requirements, and how Neuphonic can fit into your stack.
Solutions
Resources
Neuphonic © 2026
Solutions
Resources
Neuphonic © 2026
Solutions
Resources
Neuphonic © 2026