Open evaluation and benchmarking for speech models.
Public boards
Compiled August 2026. Our own harness comes next.
Speech Arena Elo. Higher is better.
| # | Model | Elo | Kind |
|---|---|---|---|
| 01 | Simba 3.2 | 1239 | API |
| 02 | Qwen-Audio-3.0-TTS-Plus | 1237 | API |
| 03 | Luna TTS | 1219 | API |
| 04 | Gemini 3.1 Flash TTS | 1212 | API |
| 05 | Sonic 3.5 | 1202 | API |
| 06 | Eleven v3 | 1177 | API |
| 07 | Fish Audio S2 Pro | 1124 | Open |
| 08 | Kokoro 82M | 1059 | Open |
Open ASR mean WER. Lower is better.
| # | Model | WER | Kind |
|---|---|---|---|
| 01 | ARK-ASR-3B | 5.04% | Open |
| 02 | Granite Speech 4.1 2B | 5.33% | Open |
| 03 | Cohere Transcribe | 5.42% | Open |
| 04 | Pulse Pro | 5.42% | API |
| 05 | Canary-Qwen-2.5B | 5.63% | Open |
| 06 | ElevenLabs Scribe v2 | 5.83% | API |
| 07 | Parakeet TDT 0.6B v2 | 6.05% | Open |
| 08 | Whisper Large v3 | 7.44% | Open |
VoxCeleb1-O equal error rate. Lower is better.
| # | Model | EER | Kind |
|---|---|---|---|
| 01 | ReDimNet2-B6 | 0.29% | Open |
| 02 | Wespeaker ResNet293 | 0.45% | Open |
| 03 | ERes2Net-Large | 0.57% | Open |
| 04 | ERes2NetV2 | 0.61% | Open |
| 05 | CAM++ | 0.73% | Open |
| 06 | ECAPA-TDNN | 0.73% | Open |
pyannote 3.1 DER, full protocol. Lower is better.
| System | Set | DER | Kind |
|---|---|---|---|
| pyannote 3.1 | AISHELL-4 | 12.2% | Open |
| pyannote 3.1 | AMI headset | 18.8% | Open |
| pyannote 3.1 | AMI distant | 22.4% | Open |
Sources: Speech Arena, Open ASR, published VoxCeleb and pyannote evals.