Artificial Analysis measured the average time to first audio of 37 speech-to-speech models on Big Bench Audio. Here are the 25 fastest, with each one's speech-reasoning score and cost beside it. The open-weights Raon SpeechChat starts speaking in 0.04 seconds, but the fast models are often the weaker reasoners: the table shows both.
Average time to first audio, Artificial Analysis, ten fastest seconds, lower is faster
Measured Artificial Analysis, Big Bench Audio question set (v1.2 harness), retrieved Oct 17, 2026.
As of All figures measured by Artificial AnalysisBig Bench Audio, 1,000 audio questions; time to first audio averaged over the question set
| Rank | Model | Lab | Time to first audio (s) | Speech reasoning (Big Bench Audio, %) | Cost per hour of input audio (USD, BBA subset) | Open weights | Figure |
|---|---|---|---|---|---|---|---|
| 1 | Raon SpeechChat1 | Krafton | 0.04 | 57.5 | — | Yes | Measured |
| 2 | Deepslate Opal1 | Deepslate | 0.44 | 85.0 | 6.48 | No | Measured |
| 3 | Gemini 2.5 Flash Native Audio Dialog1 | 0.63 | 68.6 | 1.42 | No | Measured | |
| 4 | Grok Voice 2.01 | SpaceXAI | 0.70 | 97.2 | 4.80 | No | Measured |
| 5 | Grok Voice Fast 1.01 | SpaceXAI | 0.78 | 93.3 | 3.00 | No | Measured |
| 6 | Qwen3.5 Omni Flash Realtime1 | Alibaba Cloud | 0.79 | 59.0 | 0.00 | No | Measured |
| 7 | GPT Realtime Mini (Oct '25)1 | OpenAI | 0.81 | 63.6 | 3.04 | No | Measured |
| 8 | GPT-Realtime-1.51 | OpenAI | 0.81 | 81.4 | 11.44 | No | Measured |
| 9 | Grok Voice 3.01 | SpaceXAI | 0.81 | 98.5 | 4.80 | No | Measured |
| 10 | GPT-Realtime-2.1 Mini Minimal1 | OpenAI | 0.85 | 62.7 | 4.60 | No | Measured |
| 11 | Qwen3 Omni Realtime1 | Alibaba Cloud | 0.88 | 56.8 | 2.26 | No | Measured |
| 12 | Gemini 3.1 Flash Live Minimal1 | 0.96 | 71.3 | 1.50 | No | Measured | |
| 13 | GPT-Realtime-2.1 Minimal1 | OpenAI | 0.97 | 86.8 | 11.31 | No | Measured |
| 14 | GPT Realtime (Aug '25)1 | OpenAI | 0.98 | 83.3 | 11.08 | No | Measured |
| 15 | GPT-Realtime-2 (Minimal)1 | OpenAI | 1.12 | 71.8 | 3.07 | No | Measured |
| 16 | GPT-Realtime-2 (High)1 | OpenAI | 1.14 | 96.6 | 4.14 | No | Measured |
| 17 | Nova 2.0 Sonic (Mar 2026)1 | Amazon Bedrock | 1.14 | 88.1 | 4.89 | No | Measured |
| 18 | Gemini 3.8 Live1 | 1.18 | 91.7 | 0.84 | No | Measured | |
| 19 | GPT-Realtime-2.1 High1 | OpenAI | 1.21 | 95.8 | 10.75 | No | Measured |
| 20 | GPT-Realtime-2 (Medium)1 | OpenAI | 1.22 | 93.3 | 3.97 | No | Measured |
| 21 | GPT-Live-1 (Sol, low)1 | OpenAI | 1.24 | 89.0 | 4.47 | No | Measured |
| 22 | Grok Voice 1.01 | SpaceXAI | 1.25 | 97.1 | 3.00 | No | Measured |
| 23 | GPT-4o mini Realtime (Dec 2024)1 | OpenAI | 1.27 | 68.9 | 5.75 | No | Measured |
| 24 | GPT-Live-1 (Astra, medium)1 | OpenAI | 1.34 | 90.1 | 5.83 | No | Measured |
| 25 | Gemini 3.8 Live Extended Thinking (High)1 | 1.35 | 97.7 | 3.50 | No | Measured |
Ranked by average time to first audio (the seconds until the model emits the first piece of audio), as Artificial Analysis's independent testing reports it; the table keeps the 25 fastest of the 37 models with a figure. Rows are model configurations as Artificial Analysis lists them, so one model can appear several times at different reasoning settings (for example GPT-Realtime-2 at Minimal, Medium and High). "Speech reasoning" is accuracy on Big Bench Audio, shown beside the ranked column and never blended with it; "Cost per hour of input audio" is Artificial Analysis's cost to complete a fixed 40-question subset, normalized to an hour of input audio, and it lists a cost of 0 for some models, so read a 0.00 as listed rather than as proof the model is free. Artificial Analysis lists the Grok Voice models under the host name SpaceXAI. Lab-stated prices and latencies are left out. No averages or totals are ours. Figure labels: how they work.
Artificial Analysis defines time to first audio (TTFA) as the average number of seconds a model needs to produce the first piece of audio output, measured across its Big Bench Audio question set, and calls it "a critical indicator of perceived responsiveness in voice agent applications".2 Big Bench Audio is its own benchmark of 1,000 audio questions adapted from Big Bench Hard, 250 each in formal fallacies, navigation, object counting and web of lies, spoken by 23 synthetic voices; the model gets audio in and must answer in audio.23 The accuracy column here follows the methodology page's current harness, Big Bench Audio v1.2 (May 2026 onward), which updated the judge model and grader.2 The page we read doesn't say where the requests were sent from, so absolute seconds are best compared within this table, not with latencies measured elsewhere.
Two limits apply. TTFA is a speed measure only: a model can speak sooner and answer worse, which is why the reasoning score sits beside it. And Artificial Analysis lists this as one column among several; its own headline Speech to Speech Index blends reasoning, agentic performance, arena preference and task success rather than speed.12
Krafton's open-weights Raon SpeechChat has the lowest TTFA on the leaderboard, 0.04 seconds, and Artificial Analysis names it the fastest model on its page.1 Its speech-reasoning score is 57.5%, and Artificial Analysis lists no cost figure for it, so the table is as much about the trade-off as the speed.1
Deepslate Opal starts speaking in 0.44 s while scoring 85.0% on Big Bench Audio, at a cost of $6.48 per hour of input audio on Artificial Analysis's subset.1
Google's Gemini 2.5 Flash Native Audio Dialog is third at 0.63 s, with 68.6% on Big Bench Audio.1 Its Thinking variant is far slower, 3.87 seconds, with a higher reasoning score of 90.7%, so it falls outside the top 25 table; the trade is visible within one model family.1
Grok Voice 2.0 (0.70 s, 97.2%), Grok Voice Fast 1.0 (0.78 s, 93.3%) and Grok Voice 3.0 (0.81 s, 98.5%) are all in the top ten, and Grok Voice 3.0's reasoning score is the highest of any model faster than a second.1 xAI's release notes say grok-voice-3.0 became available on October 14, 2026, at the same price as Grok Voice 2.0, with the grok-voice-latest alias moved to it and Grok Voice 2.0 deprecated.4
Alibaba Cloud's Qwen3.5 Omni Flash Realtime (0.79 s, 59.0%), GPT Realtime Mini (Oct '25) (0.81 s, 63.6%) and GPT-Realtime-1.5 (0.81 s, 81.4%) fill the middle of the top ten. The three are separated by 0.02 seconds of TTFA and 22 points of reasoning score.1
Because the leaderboard lists configurations, it shows what a reasoning setting costs in time. OpenAI's GPT-Realtime-2 is listed at 1.12 s and 71.8% at Minimal, and at 1.14 s and 96.6% at High, almost the same time for a far higher score; the Medium setting sits at 1.22 s and 93.3%.1 Google's Gemini 3.1 Flash Live goes from 0.96 s (71.3%) at Minimal to 2.99 s (96.6%) at High, while Gemini 3.8 Live and its Extended Thinking variant are listed at 1.18 s (91.7%) and 1.35 s (97.7%).1 Reasoning cost shows up as time for some families and barely at all for others, so TTFA should be read per configuration, not per brand.
Speed is not the only dimension Artificial Analysis publishes. Its page also reports conversational dynamics on a subset of Full Duplex Bench, agentic performance on τ-Voice and arena preference, and the three are not folded into this ranking.12