fetchbean

Best speech-to-speech models

Realtime voice models, scored on Big Bench Audio.

StepAudio 3 Realtime leads from StepFun, ahead of 38 others.

  1. 1StepFun

    StepAudio 3 Realtime

    1.00 big bench audio

  2. 2Alibaba

    Qwen3.5 Omni Plus Realtime

    0.99 big bench audio

  3. 3Alibaba

    Qwen Audio 3.0 Realtime Plus

    0.99 big bench audio

ranked by big bench audio · 39 models · updated Oct 13

#ModelBig Bench Audio
1StepAudio 3 Realtime
1.00
2Qwen3.5 Omni Plus Realtime
0.99
3Qwen Audio 3.0 Realtime Plus
0.99
4Step-Audio R1.1 (Realtime)
0.98
5Gemini 3.1 Flash Audio Preview (Mar 2026) - High
0.97
6Grok Voice Think Fast 1.0
0.97
7GPT-Realtime-2 High
0.97
8Grok Voice Think Fast 2.0 High
0.97
9Qwen Audio 3.0 Realtime Flash
0.96
10GPT-Realtime-2.1 High
0.96
11Grok Voice Agent
0.93
12GPT-Realtime-2 Medium
0.93
13Gemini 3.8 Live
0.92
14Gemini 2.5 Flash Native Audio Dialog Thinking
0.91
15GPT-Live-1 (Astra, medium)
0.90
16GPT-Live-1 (Sol, low)
0.89
17Nova 2.0 Sonic
0.88
18GPT-Realtime-2.1 Minimal
0.87
19Deepslate Opal
0.85
20GPT Realtime (Aug 2025)
0.83
21GPT-Realtime-1.5
0.81
22GPT-Realtime-2.1 Mini High
0.75
23GPT-Realtime-2 Minimal
0.72
24Gemini 3.1 Flash Audio Preview (Mar 2026) - Minimal
0.71
25Gemini 2.5 Flash Native Audio Dialog
0.69
26GPT-4o mini Realtime (Dec '24)
0.69
27Higgs Realtime
0.69
28GPT Realtime Mini (Oct 2025)
0.64
29GPT-Realtime-2.1 Mini Minimal
0.63
30Qwen3.5 Omni Flash Realtime
0.59
31Qwen3 Omni 30B A3B Instruct
0.59
32Raon SpeechChat
0.57
33Qwen3 Omni 30B A3B Instruct (Realtime)
0.57
34GPT 4o audio chatcompletions
0.54
35Freeze-Omni
0.33
36Nemotron 3 VoiceChat (V1)
0.27
37PersonaPlex
0.19
38FLM-Audio
0.16
39Moshi
0.04

what big bench audio means

A spoken reasoning benchmark, scored from zero to one. Higher is better.