Announcing The Artificial Analysis Speech To Speech Index
Source
artificialanalysis.ai
Author
Artificial Analysis articles
Date
Why it matters
If you build voice agents, no model completes more than about half of customer-service tasks end to end. Trade off the leaders on latency (0.44s to 2.98s to first audio) and cost.
Key takeaways · AI-distilled
The Speech to Speech Index weights three datasets equally: Big Bench Audio for reasoning, a Full Duplex Bench subset for pauses, turn-taking and interruptions, and tau-Voice for end-to-end airline, retail and telecom service tasks.
GPT-Realtime-2 (High) led the launch index at 77.2%, ahead of Grok Voice Think Fast 1.0 at 75.7%, GPT-Realtime-1.5 at 72.0% and Gemini 3.1 Flash Live Preview (High) at 69.5%.
Agentic performance on tau-Voice was the hardest dimension: Grok Voice Think Fast 1.0 led at 52.1% and every model was below 53%, while speech reasoning clustered near the top, led at 97.1%.
Speed and cost trade against score: Deepslate Opal had the fastest time to first audio at 0.44s with a 62.1% score, and Gemini 3.1 Flash Live Preview (Minimal) was cheapest at $1.50 per hour of input audio with 56.6%.
