Artificial Analysis Tracks How Fast Voice AI Actually Starts Speaking
Artificial Analysis launched Time to First Audio, a new TTS latency benchmark run daily on 12 prompts per model. Streaming models median 376 ms to first audible speech; non streaming models median 2.0 seconds, a 5x gap.

Takeaways
- Artificial Analysis launched Time to First Audio, a new TTS latency benchmark run daily on 12 prompts per model.
- Streaming models median 376 ms to first audible speech; non streaming models median 2.0 seconds, a 5x gap.
- Speechify Simba 3.5 leads at 116 ms with the top Elo under 200 ms; Inworld Realtime TTS-2 follows at 186 ms.
- ElevenLabs Eleven v4 Turbo hits 258 ms vs 653 ms for Eleven v4 at matching Arena Elo quality.
- Gradium TTS shows tightest variance at 13 ms P25 to P75; Gemini 3.1 Flash TTS swings 908 to 975 ms.
- Full leaderboard and methodology live at artificialanalysis.ai .
In voice conversations, the delay before speech begins shapes how quickly an agent can take its turn. Artificial Analysis has added Time to First Audio, or TTFA, to its text-to-speech leaderboard. Daily tests currently show a median of 376 milliseconds across 65 streaming models and 2.0 seconds across 20 non-streaming models, a 5.3-fold gap.
Conventional TTS benchmarks measure naturalness, pronunciation, throughput, and price. TTFA isolates output startup latency after the complete text has been submitted, helping developers compare endpoints and budget each stage of a voice pipeline. It excludes upstream work such as speech recognition and language-model generation.
A clock built around audibility
TTFA starts when the complete text request is submitted and stops when speech becomes audible. The calculation combines the arrival time of the first audio chunk with any silence at the beginning of the generated clip. An endpoint that returns bytes quickly still receives a slower score when those bytes contain half a second of silence.
The audibility detector uses the first 10-millisecond window whose root mean square energy exceeds −40 dBFS. Root mean square energy summarizes the signal level, while dBFS measures that level relative to the maximum available in digital audio. The published score is a seven-day rolling mean, which reduces the influence of network jitter and occasional slow requests.
One warm-up, 12 scored prompts
Every model runs once per day at a randomized time with a randomly selected English voice. An unscored warm-up request reduces cold-start effects, followed by 12 prompts that produce the recorded results.
| Characters | Prompts | Composition |
|---|---|---|
| 20 to 45 | 3 | One fixed, two changed daily |
| 60 to 120 | 3 | One fixed, two changed daily |
| 180 to 300 | 3 | One fixed, two changed daily |
| 450 to 700 | 3 | One fixed, two changed daily |
All models receive the same prompts in the same shuffled order on a given day. Keeping one prompt fixed in each band supports comparisons over time, while rotating the other two reduces dependence on a narrow test set.
Streaming endpoints stop the network clock when the first raw-audio chunk arrives, with detected leading silence added to the result. The test assumes uninterrupted playback after that point, so TTFA excludes midstream stalls. Buffered endpoints run until the complete response arrives because playback cannot begin earlier.
Streaming cuts the median wait fivefold
Across the current leaderboard, the 65 streaming models have a median TTFA of 376 milliseconds. The 20 buffered models reach 2.0 seconds, giving them roughly 5.3 times the startup latency. The result reflects endpoint behavior as well as synthesis speed: an API that waits for the full file delays playback even when generation itself is fast.
Conversational systems generally need streaming output because TTS consumes only part of the turn-time budget. A two-second buffered response already exceeds the total latency target used by many phone agents and interactive voice applications.
The fastest voices start below 300 milliseconds
In the reported snapshot, Speechify’s Simba 3.5 starts in 116 milliseconds and has the highest Arena Elo among models below 200 milliseconds. Inworld Realtime TTS-2 follows at 186 milliseconds.
| Model | Provider | TTFA |
|---|---|---|
| Simba 3.5 | Speechify | 116 ms |
| Realtime TTS-2 | Inworld | 186 ms |
| Eleven v4 Turbo | ElevenLabs | 258 ms |
| Drama 3 | Fish Audio | 258 ms |
| Sonic 3.6 | Cartesia | 287 ms |
Eleven v4 Turbo records 258 milliseconds, compared with 653 milliseconds for Eleven v4 at matching Arena quality. That comparison quantifies the latency reduction from ElevenLabs’ dedicated Turbo endpoint.
Consistency also affects turn timing. Gradium TTS has the tightest spread on the leaderboard, placing the middle 50 percent of requests within a 13-millisecond window. Simba 3.2 follows at 23 milliseconds, while Inworld Realtime TTS-2 reaches 32 milliseconds. Google DeepMind’s Gemini 3.1 Flash TTS ranges from 908 to 975 milliseconds, and Mistral’s Voxtral 3 TTS ranges from 455 to 576 milliseconds.
Budget the entire voice turn
Teams building phone agents, interactive voice response replacements, and real-time assistants often target roughly 800 milliseconds from the end of a user’s turn to the start of the agent’s reply. That budget can include several sequential steps:
- Speech endpointing and transcription
- Language-model processing and token generation
- Text buffering or chunking
- TTS startup latency
- Network transport and audio playback
TTFA measures the fourth step after a complete text input has been submitted. It does not capture time to final audio, interruptions later in a stream, or the behavior of systems that feed partial language-model output into TTS. Developers should combine it with end-to-end traces from their own application and deployment region.
Several turbo and real-time models now begin speaking within 300 milliseconds while retaining competitive human-preference scores. For audiobook generation, podcast synthesis, and offline voice-over rendering, startup latency is amortized across longer files, leaving quality, throughput, and price with greater weight.
Five axes for a practical shortlist
The leaderboard now exposes five complementary measurements, each covering a different part of model selection.
| Metric | What it measures | Preferred direction |
|---|---|---|
| Time to First Audio | Delay from complete text submission to audible speech, reported as a seven-day rolling mean | Lower |
| Arena Elo | Human preference from pairwise listening votes | Higher |
| Pronunciation Robustness | Performance on contextual disambiguation, text normalization, sequence fidelity, and standalone terms | Higher |
| Throughput | Median input characters converted to speech per second over the previous three days | Higher |
| Price | Normalized cost in US dollars per million characters | Lower |
Arena Elo comes from pairwise votes in the Artificial Analysis Text to Speech Arena. A Bradley-Terry model, fitted through maximum likelihood, converts those preferences into relative scores. Price normalization converts subscription, token, and inference-time billing so commercial APIs and hosted open-weight models can share one scale.
The current leaderboard places Eleven v4 first in overall Elo at 1,340. Amazon Polly Standard leads raw throughput at 924 characters per second through Amazon Bedrock, while Kokoro 82M v1.0 has the lowest listed price at $0.65 per million characters on Replicate. Live rankings can change as daily measurements and preference votes arrive; the full methodology defines each calculation and testing condition.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report

