Speech-to-speech models compared
A speech-to-speech model hears the caller and speaks back in one step, with no separate transcript in between. That can cut the wait, but it is priced per audio token and is harder to inspect when a call goes wrong.
Calls where the agent finished the task, measured by tau-Voice. Higher is better.
- Best
- 38%xAI Grok Voice (tested grok-voice-agent)
- Charted
- 3 of 10Measured by tau-Voice
- Weakest
- 26%Google Gemini 3.8 Live (tested Gemini Live 2.5 Flash)
- Measured by others
- 5 of 10Have an independent figure
10 compared, read 29 October 2026 or later. Click a bar to open its row, a column heading to sort, ↗ for the source. Tick Compare on up to four rows.
Showing 10 of 10
Calls where the agent finished the task, measured by tau-Voice
Measured by othersShare of retail, airline and telecom tasks completed over 8 kHz phone audio with background noise, accents and interruptions, from the open tau-Voice benchmark. Each was tested on an earlier version of the model, named on its bar. Higher is better.
Not on this chart. Not measured by tau-Voice: OpenAI GPT-Realtime-2.1 Mini, OpenAI GPT-Live 1, Amazon Nova 2.5 Sonic, Microsoft Azure Voice Live API, Ultravox Realtime, Hume AI Empathic Voice Interface, Kyutai Moshi.
- Tasks done, bestxAI Grok Voice (tested grok-voice-agent) 38%
- Tasks done, clean audio, bestxAI Grok Voice (tested grok-voice-agent) 51%
- Cost a minute, lowestOpenAI GPT-Realtime-2.1 Mini $0.0132
Bar length is the figure itself: the longest bar is best on tasks done and tasks done, clean audio, the shortest on cost a minute.
| tau-Voice, higher is better | tau-Voice, higher is better | US$ a minute | Checks✓ yes, ✕ no, ~ partly | Measured by otherssingle figures | |
|---|---|---|---|---|---|
| xAI (docs now branded SpaceXSI) Grok Voice (grok-voice-3.0) speech-to-speech API | 38%Best | 51%Best | $0.08 | Tool calls | None found |
| OpenAI GPT-Realtime-2.1 | 35% | 49% | $0.0422 | Tool calls | None found |
| Google Gemini 3.8 Live (Live API) | 26% | 31% | Not comparable | Tool calls | None found |
| OpenAI GPT-Realtime-2.1 Mini | Not measured by tau-Voice | Not measured by tau-Voice | $0.0132Lowest | Tool calls | None found |
| OpenAI GPT-Live 1 | Not measured by tau-Voice | Not measured by tau-Voice | $0.05 | Tool calls | None found |
| Amazon (AWS) Nova 2.5 Sonic | Not measured by tau-Voice | Not measured by tau-Voice | Not comparable | British English Tool calls | None found |
| Microsoft Azure Voice Live API | Not measured by tau-Voice | Not measured by tau-Voice | Not comparable | British English Tool calls | None found |
| Ultravox (Fixie.ai, acquired by Inworld) Ultravox Realtime (open-weight speech model plus hosting)Open weights | Not measured by tau-Voice | Not measured by tau-Voice | $0.05 | Tool calls | |
| Hume AI Empathic Voice Interface (EVI) | Not measured by tau-Voice | Not measured by tau-Voice | Not published | Not published | None found |
| Kyutai MoshiOpen weights | Not measured by tau-Voice | Not measured by tau-Voice | Not published | Not published | |
Vendor claims, data regions and compliance are in each row’s Details, in the vendor’s own words. Claims are never charted or ranked. RingTru has measured none of these yet. Add up a whole call in the calculator.
How to read these numbers
- Audio tokens
- These models bill per audio token, not per minute. A minute of audio is a different number of tokens for each vendor, so the arena converts to minutes only where the vendor publishes the rate.
- No transcript in the middle
- A chained agent keeps a transcript you can read when a call fails. A speech-to-speech model may not, which makes testing and evidence harder. Our guide on speech-to-speech and chained agents covers the trade-off.
Go deeper
Your agent, tested on real calls.
- test calls
- 506
- passed
- 96.8%
- median reply
- 0.94 s
Holds the new day when a caller changes their mind.