Which TTS model sounds most human?
Start with Audio Realism. Its blind pairwise listening protocol covers selected American-English voices across phone-agent, conversational, and explainer prompts.
Voice systems
Voice benchmarks, separated by what they measure
There is no single defensible voice-agent score yet. These tables keep spoken reasoning, workflow success, conversation dynamics, experience, and latency in their original protocols so the numbers remain interpretable.
Source snapshots refreshed
Five measurement lanesVOICE / S2S
Reasoning, task completion, conversation dynamics, experience, and latency stay separate.
01
Spoken reasoning
Does the system understand and answer spoken requests?
05
Task completion
Can it follow policy, use tools, and finish a workflow?
06
Conversation dynamics
Can it handle turns, interruptions, ambiguity, and state?
02
Voice experience
Is the exchange natural, robust, and responsive?
01
Latency
How long do responses, tool calls, and full tasks take?
Current source-backed results
The tabs are separate leaderboards, not ingredients in a BenchLM ranking. Audio Realism uses blind listening preferences; VoiceBench uses its own overall score; AudioAgentBench shows checks passed; Full-Duplex-Bench reports strict pass@1 and seconds.
| Rank | Model | Elo-style rating | BT uncertainty | Win rate | Battles | Avg. generation |
|---|---|---|---|---|---|---|
| 1 | Bland Speech v3 Bland AI | 1365 | ±23.9 | 82.5% | 372 | 1.68s |
| 2 | MAI-Voice-2 Microsoft | 1224 | ±19.8 | 68.9% | 376 | 1.50s |
| 3 | Grok TTS xAI | 1156 | ±18.8 | 60.9% | 378 | 2.27s |
| 4 | Gemini 2.5 Pro TTS Preview Google | 1109 | ±18.2 | 54.9% | 390 | 6.95s |
| 5 | Cartesia Sonic 3.5 Cartesia | 1096 | ±18.2 | 51.6% | 395 | 1.55s |
| 6 | MiniMax Speech-02 HD MiniMax | 1093 | ±18.6 | 52.4% | 380 | 3.36s |
| 7 | Gemini 3.1 Flash TTS Preview Google | 1037 | ±18.4 | 41.5% | 388 | 4.99s |
| 8 | GPT-4o mini TTS OpenAI | 1011 | ±18.9 | 42.8% | 362 | 2.01s |
| 9 | Gemini 2.5 Flash TTS Preview Google | 985 | ±18.9 | 37.9% | 385 | 5.16s |
| 10 | Lightning v3.1 Pro Smallest AI | 952 | ±19.4 | 34.0% | 380 | 1.94s |
| 11 | ElevenLabs Eleven v3 ElevenLabs | 938 | ±20.1 | 29.1% | 364 | 3.35s |
| 12 | Inworld TTS-1.5 Max Inworld | 763 | ±24.9 | 14.2% | 373 | 4.97s |
Snapshot sources: benchmark-owner leaderboard and run records. Full-Duplex-Bench v3 values are transcribed from the official paper table.
Benchmark coverage
| Benchmark | Spoken reasoning | Task completion | Conversation dynamics | Voice experience | Latency | Evidence |
|---|---|---|---|---|---|---|
| Audio Realism Benchmark Blind listening comparisons measure how human text-to-speech output sounds across phone-agent, conversational, and explainer speech. | Live snapshot Machine-readable public leaderboard | |||||
| VoiceBench A spoken question-answering suite spanning open-ended, knowledge, reasoning, instruction-following, and safety tasks. | Live snapshot Machine-readable public leaderboard | |||||
| AudioAgentBench Evaluates real-time voice agents on policy compliance, tool use, grounding, ambiguity, and state tracking. | Live snapshot Machine-readable public run records | |||||
| Full-Duplex-Bench v3 Tests full-duplex voice agents on multi-tool tasks with human speech, disfluencies, interruptions, and latency measurements. | Paper results Results transcribed from the official paper |
What the current results can answer
Which model answers spoken questions best?
Start with VoiceBench. Check its architecture and individual task columns before relying on the owner-defined overall score.
Which agent completes service workflows?
Use AudioAgentBench for recent public run aggregates, then compare human and synthesized audio rather than merging them.
Which full-duplex system balances success and speed?
Use Full-Duplex-Bench v3. Strict pass@1 and latency answer different questions and should be inspected together.