Artificial Analysis Measures TTS Start Time: Latency Ranges From 76 ms to Nearly 1 Second
Artificial Analysis has added a Time to First Audio metric to its text-to-speech (TTS) ranking: the time until the first audible sound. Picture a voice bot: you ask a question, it stays silent for a second, and it feels like the connection dropped. For leading models, that pause ranges from 76 ms to nearly a second. For live conversation, choose streaming models.
AI-processed from X: @ArtificialAnlys; edited by Hamidun News
Say you are building a voice agent for customer support. The user asks a question and the language model has already written an answer, but the line is silent: the speech synthesizer hasn't produced a single sound yet. At that moment the person isn't thinking about the quality of the intonation. They decide the connection has dropped. Artificial Analysis set out to measure exactly this pause on its own. It added a Time to First Audio metric to its speech synthesis benchmarks (Text to Speech, TTS, which turns text into voice). For leading models, the wait falls between 76 ms and nearly a second.
The TTS comparison already covered voice quality, pronunciation and generation speed. But none of those numbers answers a simple question: how long does a person wait before hearing anything? That wait determines whether an agent feels like a live conversation partner or an answering machine with pauses.
How the measurement works
The methodology is fairly strict. Each model is tested daily on the same 12 prompts, with texts of four different lengths. The clock stops not at the first byte of data but when audible speech appears in the response. The final ranking shows a 7-day moving average, so a single random load spike doesn't distort the picture.
For conversation, the gap between the two ends of the range is huge: a person barely notices a hundred milliseconds, while a full second already feels like a stumble.
Who is fastest
Among the frontier models, Simba 3.5 showed the best balance. It starts speaking in 116 ms and has the highest Arena Elo of all models that need less than 200 ms. Arena Elo is a voice quality rating built from head-to-head comparisons between models. Realtime TTS-2 follows with a latency of 186 ms.
The "turbo" versions look interesting. Eleven v4 Turbo starts in 258 ms, while the standard Eleven v4 takes 653 ms, and their Arena Elo is the same. In other words, a speedup of roughly two and a half times came without a loss of quality in this ranking. Fish Audio Drama 3 shows the same 258 ms, and Sonic 3.6 takes 287 ms.
Streaming as the dividing line
The main split is not between brands but between architectures. 65 models can stream: they send audio in chunks while the rest of the text is still being processed. Their median start time is 376 ms. Another 20 models do not stream: the user gets nothing until the entire phrase has been generated. Their median is 2.0 seconds. The longer the utterance, the worse this approach performs, since you have to wait for the whole generation to finish.
What to do about it
The practical takeaway for anyone building voice agents, call-center scenarios or voice-over in apps is fairly straightforward:
- Don't choose a TTS model by its quality ranking alone. A model with an excellent voice but a one-second pause can kill the feeling of a live dialogue.
- For conversational scenarios, pick streaming models and look at turbo variants: Eleven v4 Turbo has the same quality score as the standard version but starts 2.5 times faster.
- Remember that an agent's total latency is the sum of speech recognition, the language model's response and synthesis. TTS is only the last link, but it determines when the user first hears the answer.
Limitations
These are measurements from a single independent service on 12 prompts, and in a real product the numbers will differ: network, region, load and the length of your utterances will add tens or hundreds of milliseconds of their own. In addition, the announcement does not disclose every detail, for example how exactly the four input lengths were chosen. So treat the ranking as a guide for building a shortlist, and make the final decision after your own test on your own traffic. Still, the fact that the industry now has a shared public measurement of time to first audio helps compare models on what the user actually feels.
- Nov 21, 2026VibeCBS collects project nicknames in advance: the agent finds 2.8 times more documents
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026Sakana AI Opens Applied Research Lab for Market and Economic Modeling
- Nov 21, 2026Meta Is Reportedly Preparing the $300 Muse Charm AI Tamagotchi With a Year of Verizon 5G
Key Questions Answered
How the measurement works?
The methodology is fairly strict. Each model is tested daily on the same 12 prompts, with texts of four different lengths. The clock stops not at the first byte of data but when audible speech appears in the response. The final ranking shows a 7-day moving average, so a single random load spike doesn't distort the picture. For conversation, the gap between the two ends of the range is huge: a person barely notices a hundred milliseconds, while a full second already feels like a stumble.
Who is fastest?
Among the frontier models, Simba 3.5 showed the best balance. It starts speaking in 116 ms and has the highest Arena Elo of all models that need less than 200 ms. Arena Elo is a voice quality rating built from head-to-head comparisons between models. Realtime TTS-2 follows with a latency of 186 ms. The "turbo" versions look interesting. Eleven v4 Turbo starts in 258 ms, while the standard Eleven v4 takes 653 ms, and their Arena Elo is the same.
What are the key details on Streaming as the dividing line?
The main split is not between brands but between architectures. 65 models can stream: they send audio in chunks while the rest of the text is still being processed. Their median start time is 376 ms. Another 20 models do not stream: the user gets nothing until the entire phrase has been generated. Their median is 2.0 seconds. The longer the utterance, the worse this approach performs, since you have to wait for the whole generation to finish.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.