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
No single score covers every voice-agent workload. These tables keep reasoning, workflow success, conversation dynamics, experience, and latency in their own protocols.
Source snapshots refreshed
Five measurement lanesVOICE / S2S
Reasoning, task completion, conversation dynamics, experience, and latency stay separate.
08
Spoken reasoning
Does the system understand and answer spoken requests?
11
Task completion
Can it follow policy, use tools, and finish a workflow?
17
Conversation dynamics
Can it handle turns, interruptions, ambiguity, and state?
35
Voice experience
Is the exchange natural, robust, and responsive?
23
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 | Freya TTS Adam V1 Freya | 1442 | ±23.8 | 79.3% | 381 | 2.41s |
| 2 | Kalpa TTS Beta v0.1 Kalpa Labs | 1380 | ±18.8 | 76.3% | 586 | 2.91s |
| 3 | Bland Speech v3 Bland AI | 1379 | ±18.4 | 77.8% | 626 | 1.40s |
| 4 | Cartesia Sonic 3.6 Cartesia | 1341 | ±27.4 | 73.6% | 250 | 1.82s |
| 5 | Deepgram Flux TTS deepgram | 1310 | ±29.7 | 70.2% | 208 | 7.21s |
| 6 | ElevenLabs Eleven v3 Conversational ElevenLabs | 1238 | ±17.7 | 64.6% | 508 | 2.61s |
| 7 | Murf Falcon 2 Murf AI | 1232 | ±27.6 | 63.3% | 218 | 1.62s |
| 8 | Cartesia Sonic 3.5 Cartesia | 1218 | ±13.9 | 63.1% | 842 | 1.61s |
| 9 | MAI-Voice-2 Microsoft | 1208 | ±15.8 | 61.1% | 640 | 1.40s |
| 10 | HeyGen Voice heygen | 1185 | ±21.7 | 62.2% | 304 | 14.96s |
| 11 | Grok TTS xAI | 1179 | ±15.2 | 58.3% | 667 | 1.79s |
| 12 | MAI-Voice-2-Flash microsoft | 1137 | ±31.0 | 51.2% | 166 | 1.65s |
| 13 | MiniMax Speech 2.8 HD MiniMax | 1135 | ±26.6 | 51.1% | 217 | 7.08s |
| 14 | MiniMax Speech-02 HD MiniMax | 1121 | ±15.0 | 50.9% | 674 | 2.95s |
| 15 | Gemini 2.5 Pro TTS Preview Google | 1110 | ±14.7 | 49.6% | 688 | 5.96s |
| 16 | GPT Realtime 2 OpenAI | 1103 | ±19.9 | 47.2% | 362 | 2.46s |
| 17 | Google Gemini 3.8 Flash TTS google | 1080 | ±18.1 | 48.2% | 423 | 5.01s |
| 18 | GPT Realtime 2.1 OpenAI | 1064 | ±24.5 | 46.2% | 236 | 3.21s |
| 19 | Eleven v4 elevenlabs | 1031 | ±21.4 | 42.1% | 304 | 2.53s |
| 20 | GPT-4o mini TTS OpenAI | 1029 | ±15.3 | 40.3% | 650 | 1.99s |
| 21 | Gemini 3.1 Flash TTS Preview Google | 1025 | ±15.1 | 37.2% | 683 | 4.20s |
| 22 | Voxtral TTS 3 mistral | 1019 | ±21.4 | 41.3% | 300 | 2.26s |
| 23 | Eleven v4 Turbo elevenlabs | 998 | ±20.7 | 38.3% | 332 | 1.90s |
| 24 | ElevenLabs Eleven v3 ElevenLabs | 970 | ±15.7 | 30.5% | 673 | 2.83s |
| 25 | Gemini 2.5 Flash TTS Preview Google | 970 | ±15.4 | 32.5% | 690 | 4.32s |
| 26 | Fish Audio Drama 3 fish-audio | 958 | ±21.1 | 34.2% | 330 | 2.01s |
| 27 | Lightning v3.1 Pro Smallest AI | 955 | ±15.8 | 31.0% | 675 | 1.73s |
| 28 | Fish Audio S2.1 Pro Fish Audio | 953 | ±20.4 | 33.2% | 373 | 2.77s |
| 29 | Magpie TTS Multilingual nvidia | 875 | ±17.5 | 24.3% | 588 | 15.20s |
| 30 | Inworld Realtime TTS-2 Inworld | 860 | ±28.6 | 24.7% | 219 | 2.56s |
| 31 | Qwen-Audio-3.0 TTS Flash alibaba | 854 | ±33.3 | 23.0% | 174 | 2.76s |
| 32 | Inworld TTS-1.5 Max Inworld | 789 | ±19.3 | 15.1% | 667 | 4.11s |
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 |
|---|---|---|---|---|---|---|
| Dessa launch recordings Block shares live recordings of people talking with Dessa, a full-duplex model trained only on human speech. Block publishes no benchmark scores; the model is waitlist-only and coming soon. | Provider results Launch page with 6 featured and 15 additional recordings, October 2026 | |||||
| Eleven Realtime provider evaluation ElevenLabs reports separate speech-reasoning, conversation, human-preference, task-success, and response-latency measurements for Eleven Realtime v1 Mini and Max. | Provider results Provider-published selected figures; no shared overall score | |||||
| Voxtral 3 launch evaluation Mistral reports 4.0% FLEURS word error rate and 18.4% diarization cpWER for Voxtral Small Transcribe 3, the lowest cpWER of the seven systems in its chart. | Provider results Exact values in the chart data of the October 20, 2026 launch post | |||||
| Ming-3.0 speech evaluation inclusionAI reports 0.800% Chinese and 1.584% English word error rate on the Seed-TTS testset for Ming-3.0-tts-0.5B, and 3.45 Chinese CER with 3.69 English WER for Ming-3.0-audio-flash. | Provider results Exact values in the Hugging Face model cards; the audio and ASR cards publish them as table images |
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.