Speech to Speech AI Model & Provider Leaderboard

Name: AA-Speech to Speech Index
Creator: Artificial Analysis
License: https://artificialanalysis.ai/docs/legal/Terms-of-Use.pdf

Analysis and comparison of Speech to Speech models & API providers. Artificial Analysis has analyzed speech to speech models and hosting providers across different characteristics including their reasoning quality, conversational dynamics, generation time and price.

For further details, see the methodology page.

Text to Speech Leaderboard

Speech to Text Leaderboard

Speech Agent Arena Leaderboard

Speech Agent Arena Overview

AA-Speech to Speech Index

Weighted average of Speech Reasoning, Agentic Performance, Arena Preference, and Task Success Rate results · Higher is better

Speed

Time to first audio (seconds) on Big Bench Audio · Lower is better

Cost per Hour of Input Audio

Cost to complete fixed 40-question Big Bench Audio subset based on length of input audio, normalized to hourly basis · Lower is better

Artificial Analysis Speech to Speech Index

Weighted average of Speech Reasoning, Agentic Performance, Arena Preference, and Task Success Rate results · Only models with all data available shown · Higher is better

Equal-weighted score across Speech Reasoning (Big Bench Audio), Agentic Performance (𝜏-Voice), Arena Preference (Arena Score), and Task Success Rate. Models must have results for all four components to receive an index score.

Preference Elo is converted into expected preference against an 800-Elo baseline: Arena Score = 100 / (1 + 10^((800 − frozen Elo) / 400)). The rating is frozen when a model first becomes eligible for publication, so it does not track the live Elo shown on the Arena charts.

Speech Reasoning

Speech Reasoning (Big Bench Audio)

Speech reasoning: based on the Artificial Analysis Big Bench Audio dataset · Higher is better

Big Bench Audio is an Artificial Analysis benchmark comprising 1,000 audio questions adapted from Big Bench Hard, designed to test the reasoning ability of native audio models. Questions span four categories (250 each): Formal Fallacies (determine whether an argument is deductively valid or invalid), Navigate (determine whether navigation steps return to the starting point), Object Counting (count the number of a specific item class from a list of possessions), and Web of Lies (evaluate the truth value of a Boolean function expressed as a word problem). Models receive an audio input and generate an audio output, which is evaluated as correct or incorrect.

Agentic Performance

Agentic Performance (𝜏-Voice)

Proportion of replica customer service scenarios resolved while acting as a customer support agent, based on the 𝜏-Voice benchmark · Higher is better · Only full duplex models

𝜏-Voice measures task completion in real-world replica customer service scenarios. A voice user simulator calls the tested model with a specific domain problem (e.g., requesting a flight change, disputing a retail charge, resolving a telecom issue) across domains inherited from 𝜏-bench and 𝜏²-bench. The tested model is prompted as a customer support agent with access to domain-specific tools and a policy document. Each task has a single valid database end-state; evaluation compares final state against this ground truth. Charts use average performance across 3 trials where available. For more detail on the original benchmark see the 𝜏-Voice paper.

Note: Following models based on 1 trial: GPT-Live-1 (Astra, medium), OpenAI, GPT-Live-1 (Sol, low), OpenAI, GPT-Realtime-2 (Minimal), OpenAI; Following models based on 2 trials: GPT-Realtime-2.1 High, OpenAI

Agentic Performance (𝜏-Voice) by Domain

Proportion of replica customer service scenarios resolved by domain, based on the 𝜏-Voice benchmark · Higher is better · Only full duplex models

Speech Agent Arena

Arena Preference Elo

Preference Elo measures human participants' overall preference after they complete the same scenario with two unidentified Speech to Speech models, such as booking a dental appointment · Only includes models with tool calling · Higher is better

Pairwise preference Elo fitted with Bradley–Terry maximum likelihood, with GPT Realtime 1.5 anchored at 1000.

A Default Cascaded System is a provider’s standard voice-agent pipeline: its orchestration harness connects a speech-to-text model, an LLM, and a text-to-speech model. These entries use the provider’s default stack, rather than a custom pipeline optimized by Artificial Analysis.

Tool registration indicates whether tool definitions are pre-registered or supplied during session setup. As of July 2026:

Model	Harness	STT	LLM	TTS	Tool registration
Deepgram Voice Agent	Deepgram Voice Agent API	Nova-3	GPT-4o Mini	Aura-2	Session-level
ElevenLabs Agents	ElevenLabs Agents	Scribe v2 Realtime	Gemini 2.5 Flash	Eleven Flash v2	Pre-registered
Cartesia Line	Cartesia Line	Ink	Gemini 2.5 Flash	Sonic	Pre-registered schemas; Arena execution
Inworld Realtime	Inworld Realtime API	Inworld STT 1	Gemini 2.5 Flash	Inworld TTS 1.5 Mini	Session-level

Task Success Rate

Proportion of eligible conversations with the correct final task-completing tool call or calls · Higher is better

Percentage of eligible conversations with the correct final task-completing tool call or calls.

Tool registration indicates whether tool definitions are pre-registered or supplied during session setup. As of July 2026:

Model	Harness	STT	LLM	TTS	Tool registration
Deepgram Voice Agent	Deepgram Voice Agent API	Nova-3	GPT-4o Mini	Aura-2	Session-level
ElevenLabs Agents	ElevenLabs Agents	Scribe v2 Realtime	Gemini 2.5 Flash	Eleven Flash v2	Pre-registered
Cartesia Line	Cartesia Line	Ink	Gemini 2.5 Flash	Sonic	Pre-registered schemas; Arena execution
Inworld Realtime	Inworld Realtime API	Inworld STT 1	Gemini 2.5 Flash	Inworld TTS 1.5 Mini	Session-level

Conversational Dynamics

Conversational Dynamics (Full Duplex Bench subset)

Weighted average of pause handling, turn-taking, interruption handling, and backchannel handling from Full Duplex Bench v1 and v1.5 · Higher is better

Metrics from Full Duplex Bench v1 include Pause Handling (% of samples model correctly does not interrupt) and Turn Taking (% of samples correctly takes the turn in conversation). Metrics from Full Duplex Bench v1.5 include User Interruption Handling (% of samples where the model correctly addresses user's interruption) and Backchannel Handling (% of samples where the model correctly continues when a backchannel such as "yeah" or "alright" is played).

Conversational Dynamics (Full Duplex Bench subset) - Category Breakdown

Individual category scores ordered by conversational dynamics score · Based on subset of Full Duplex Bench v1 and v1.5

API Benchmarks

Cost per Hour of Input Audio (Big Bench Audio subset)

Cost to complete fixed 40-question Big Bench Audio subset based on length of input audio, normalized to hourly basis · Lower is better

Cost to complete a fixed 40-question Big Bench Audio subset, normalized to an hourly basis using input audio duration. It incorporates audio input, audio output, text input, text output, and separately exposed reasoning/thinking tokens. It does not include cached-token discounts or tool-call costs.

Speech Reasoning (Big Bench Audio) vs. Cost per Hour of Input Audio

Big Bench Audio score vs. cost per hour of input audio measured on fixed 40-question subset · Higher speech reasoning and lower cost are better

Most attractive quadrant

Time to First Audio

Average time to first audio (seconds) on Big Bench Audio dataset · Lower is better

Number of seconds required to generate the first token of audio output. A lower value represents a faster generation.

Speech Reasoning (Big Bench Audio) vs. Speed

Big Bench Audio score vs. time to first audio · Higher speech reasoning and lower latency are better

Most attractive quadrant

Number of seconds required to generate the first token of audio output. A lower value represents a faster generation.

Summary of Key Metrics & Further Information


Alibaba Cloud	Qwen Audio 3.0 Realtime Plus	69.5	99%	98.4%	54.6%	798	77.8%	1.54	4.42	0.03	0.18
Alibaba Cloud	Qwen3.5 Omni Plus Realtime	-	99%	-	-	843	62.0%	2.64	0.00	0.16	0.82
Alibaba Cloud	Qwen Audio 3.0 Realtime Flash	67.7	96%	96.9%	35.9%	858	81.7%	1.55	4.77	-	-
Alibaba Cloud	Qwen3.5 Omni Flash Realtime	-	59%	-	-	832	29.1%	0.79	0.00	0.16	0.82
Alibaba Cloud	Qwen3 Omni Flash	-	59%	72.7%	-	-	-	4.82	1.77	0.61	1.36
Alibaba Cloud	Qwen3 Omni Realtime	-	57%	-	-	-	-	0.88	2.26	0.65	0.99
Amazon Bedrock	Nova 2.0 Sonic (Mar 2026)	-	88%	-	-	964	57.1%	1.14	4.89	0.27	1.08
Boson AI	Higgs Realtime	-	69%	92.8%	18.6%	-	-	1.47	-	-	-
Cofe AI	FLM-Audio	-	16%	62.0%	-	-	-	-	-	-	-
Deepslate	Deepslate Opal	-	85%	85.7%	17.5%	-	-	0.44	6.48	-	-
ElevenLabs	Eleven Realtime v1 Max	77.3	99%	94.1%	47.6%	990	87.1%	0.57	4.80	-	-
ElevenLabs	Eleven Realtime v1 Mini	64.4	76%	95.8%	46.8%	981	60.3%	0.62	4.80	-	-
Google	Gemini 3.8 Live Extended Thinking (High)	82.6	98%	91.9%	68.6%	966	89.1%	1.35	3.50	-	-
Google	Gemini 3.1 Flash Live High	72.0	97%	74.3%	37.7%	1062	71.8%	2.99	1.75	0.35	1.38
Google	Gemini 3.8 Live	74.5	92%	96.1%	30.1%	1074	93.2%	1.18	0.84	-	-
Google	Gemini 2.5 Flash Native Audio Dialog Thinking	-	91%	-	-	-	-	3.87	-	0.35	1.38
Google	Gemini 3.1 Flash Live Minimal	64.2	71%	72.3%	26.2%	1089	74.6%	0.96	1.50	0.35	1.38
Google	Gemini 2.5 Flash Native Audio Dialog	-	69%	-	-	-	-	0.63	1.42	0.35	1.38
Google	Gemini 2.5 Flash Native Audio Preview (Dec 2025)	-	-	44.0%	22.8%	-	-	-	-	-	-
Google	Gemini 2.5 Flash Native Audio Preview (Sep 2025)	-	-	30.3%	-	-	-	-	-	-	-
Krafton	Raon SpeechChat	-	58%	86.2%	-	-	-	0.04	-	-	-
Kyutai	Moshi	-	4%	61.0%	-	-	-	-	-	-	-
NVIDIA	Nemotron Voicechat	-	27%	52.9%	-	-	-	-	-	-	-
NVIDIA	PersonaPlex	-	19%	91.0%	-	-	-	-	-	-	-
OpenAI	GPT-Realtime-2 (High)	73.6	97%	95.3%	39.8%	920	89.8%	1.14	4.14	1.15	4.61
OpenAI	GPT-Realtime-2.1 High	75.2	96%	95.7%	45.7%	931	91.5%	1.21	10.75	-	-
OpenAI	GPT-Realtime-2 (Medium)	-	93%	95.2%	37.4%	-	-	1.22	3.97	1.15	4.61
OpenAI	GPT-Live-1 (Astra, medium)	83.2	90%	94.9%	74.5%	1027	87.4%	1.34	5.83	-	-
OpenAI	GPT-Live-1 (Sol, low)	80.1	89%	97.3%	59.3%	1037	90.9%	1.24	4.47	-	-
OpenAI	GPT-Realtime-2.1 Minimal	70.7	87%	92.7%	38.0%	933	89.4%	0.97	11.31	-	-
OpenAI	GPT Realtime (Aug '25)	69.1	83%	93.9%	30.4%	967	89.4%	0.98	11.08	1.15	4.61
OpenAI	GPT-Realtime-1.5	70.3	81%	95.7%	38.8%	1000	85.1%	0.81	11.44	1.15	4.61
OpenAI	GPT-Realtime-2.1 Mini High	-	75%	91.7%	29.4%	-	-	4.28	3.45	-	-
OpenAI	GPT-Realtime-2 (Minimal)	63.3	72%	96.1%	30.8%	918	84.7%	1.12	3.07	1.15	4.61
OpenAI	GPT-4o mini Realtime (Dec 2024)	-	69%	-	-	-	-	1.27	5.75	0.36	1.44
OpenAI	GPT Realtime Mini (Oct '25)	56.1	64%	95.7%	15.1%	922	79.6%	0.81	3.04	0.36	1.44
OpenAI	GPT-Realtime-2.1 Mini Minimal	54.4	63%	91.8%	22.5%	840	76.7%	0.85	4.60	-	-
OpenAI	GPT-4o audio chatcompletions	-	54%	-	-	-	-	3.38	0.00	3.60	14.40
OpenAI	GPT-4o Realtime (Dec 2024)	-	-	89.8%	27.9%	-	-	1.49	2.04	1.44	5.76
Phonic	Phonic 1 (High)	87.1	99%	-	79.9%	1012	92.3%	0.61	4.20	-	-
SpaceXAI	Grok Voice 3.0	86.9	99%	97.0%	71.6%	1059	94.8%	0.81	4.80	-	-
SpaceXAI	Grok Voice 2.0	81.3	97%	95.1%	56.5%	991	94.6%	0.70	4.80	-	-
SpaceXAI	Grok Voice 1.0	73.7	97%	77.8%	52.1%	903	80.7%	1.25	3.00	-	-
SpaceXAI	Grok Voice Fast 1.0	-	93%	71.6%	27.4%	-	-	0.78	3.00	3.00	3.00
StepFun	StepAudio 3 Realtime	-	100%	98.9%	-	-	-	8.83	-	-	-
StepFun	Step-Audio R1.1 (Realtime)	-	98%	-	-	-	-	1.53	0.00	0.06	1.69
VITA	Freeze-Omni	-	33%	58.7%	-	-	-	-	-	-	-

Frequently Asked Questions

StepAudio 3 Realtime leads with a speech reasoning score of 99.7% on the Big Bench Audio dataset across 44 models evaluated. Qwen Audio 3.0 Realtime Plus follows at 99.2% and Phonic 1 (High) at 98.9%.

StepAudio 3 Realtime leads with a conversational dynamics score of 98.9% on the Full Duplex Bench dataset across 37 models evaluated. Qwen Audio 3.0 Realtime Plus follows at 98.4% and GPT-Live-1 (Sol, low) at 97.3%.

The top speech to speech models by reasoning quality are: 1. StepAudio 3 Realtime (99.7%), 2. Qwen Audio 3.0 Realtime Plus (99.2%), 3. Phonic 1 (High) (98.9%), 4. Eleven Realtime v1 Max (98.8%), and 5. Qwen3.5 Omni Plus Realtime (98.7%). Scores are based on the Big Bench Audio benchmark.

The top speech to speech models by conversational dynamics are: 1. StepAudio 3 Realtime (98.9%), 2. Qwen Audio 3.0 Realtime Plus (98.4%), 3. GPT-Live-1 (Sol, low) (97.3%), 4. Grok Voice 3.0 (97.0%), and 5. Qwen Audio 3.0 Realtime Flash (96.9%). Scores are based on the Full Duplex Bench benchmark.

Raon SpeechChat has the lowest Time to First Audio at 0.04s, followed by Deepslate Opal (0.44s) and Eleven Realtime v1 Max (0.57s). Lower values mean faster initial response.

Qwen Audio 3.0 Realtime Plus is the most affordable at $0.0331/hour for input audio, followed by Step-Audio R1.1 (Realtime) ($0.064/hour) and Qwen3.5 Omni Plus Realtime ($0.162/hour).

Full Duplex Bench evaluates how well speech to speech models handle real conversational behaviors: knowing when to speak, when to stay silent during pauses, how to respond to interruptions, and how to recognize backchannels like "yeah" or "mm-hmm". Artificial Analysis implements a subset of Full Duplex Bench v1 and v1.5.

Speech reasoning (Big Bench Audio) measures whether a model can understand and correctly answer reasoning questions delivered as audio. Conversational dynamics (Full Duplex Bench) measures whether a model can handle natural conversation flow — turn-taking, pauses, and interruptions. A model may excel at one but not the other.

Real-time conversation requires both strong conversational dynamics and low latency. By conversational dynamics score, StepAudio 3 Realtime (98.9%), Qwen Audio 3.0 Realtime Plus (98.4%), and GPT-Live-1 (Sol, low) (97.3%) lead the field. By Time to First Audio, Raon SpeechChat (0.04s), Deepslate Opal (0.44s), and Eleven Realtime v1 Max (0.57s) are the fastest. The best choice depends on whether natural conversation flow or response speed is more critical for your use case.

Benchmarks are updated regularly as new models and providers are added. Performance metrics are continuously monitored to reflect current provider capabilities and pricing changes.

Speech to Speech AI Model & Provider Leaderboard

Related Links

AA-Speech to Speech Index