Which speech model hears real people best, and fastest?
| # | Model | out of 100 | lower is better | after you stop talking | 1 in 10 is slower | Licence | ||
|---|---|---|---|---|---|---|---|---|
| 1 | Underdog Bark 3.0 ProConway Research | 100 | 1.1% | 0.05 s | 0.14 s | 100% | 1.3 GB | CC-BY-NC-4.0 |
| 2 | Underdog Bark 3.0 PlusConway Research | 100 | 0.9% | 0.04 s | 0.07 s | 100% | 0.2 GB | CC-BY-NC-4.0 |
| 3 | Parakeet TDT 0.6B v3NVIDIA | 98 | 0.9% | 0.06 s | 0.08 s | 100% | 0.4 GB | CC-BY-4.0 |
| 4 | Phonon-2Fermion Research | 98 | 1.2% | 0.06 s | 0.08 s | 100% | 0.7 GB | CC-BY-4.0 (GGUF) / Apache-2.0 (package) |
| 5 | parakeet-ultramoondream | 98 | 0.7% | 0.06 s | 0.08 s | 100% | 0.4 GB | CC-BY-4.0 |
| 6 | parakeet-redux (ternary encoder)moondream | 98 | 1.9% | 0.06 s | 0.08 s | 100% | 0.4 GB | CC-BY-4.0 |
| 7 | LFM2.5-Audio 1.5BLiquid AI | 98 | 1.0% | 0.07 s | 0.14 s | 100% | 1.5 GB | LFM Open |
| 8 | Qwen3-ASR 0.6BQwen | 96 | 0.9% | 0.08 s | 0.16 s | 100% | 1.0 GB | Apache-2.0 |
| 9 | Cohere Transcribe 03-2026Cohere | 96 | 1.5% | 0.09 s | 0.12 s | 100% | 1.3 GB | Apache-2.0 |
| 10 | Granite Speech 5.0 2BIBM | 93 | 0.9% | 0.10 s | 0.23 s | 100% | 3.1 GB | Apache-2.0 |
| 11 | Qwen3-ASR 1.7BQwen | 92 | 0.7% | 0.11 s | 0.22 s | 100% | 2.5 GB | Apache-2.0 |
| 12 | Higgs Audio v3 STTBoson AI | 91 | 0.7% | 0.11 s | 0.27 s | 100% | 2.4 GB | Apache-2.0 |
| 13 | Canary-Qwen 2.5BNVIDIA | 88 | 1.0% | 0.16 s | 0.29 s | 100% | 3.7 GB | CC-BY-4.0 |
| 14 | Whisper large-v3 turboOpenAI | 86 | 1.0% | 0.23 s | 0.26 s | 100% | 1.6 GB | MIT |
| 15 | Musci-ASR (Qwen3-ASR fine-tune)Musci Research | 85 | 0.9% | 0.18 s | 0.45 s | 100% | – | Apache-2.0 |
| 16 | Nemotron 3.5 ASR Streaming 0.6BNVIDIA | 79 | 2.9% | 0.13 s | 0.24 s | 100% | 1.3 GB | OpenMDW-1.1 |
| 17 | LFM3-Audio 2BLiquid AI | 72 | 1.4% | 0.61 s | 1.08 s | 85% | 2.1 GB | LFM Open |
| 18 | Moonshine Streaming MediumUseful Sensors | 54 | 2.7% | 1.05 s | 3.46 s | 47% | 0.2 GB | MIT |
| 19 | Voxtral Mini 4B RealtimeMistral | 40 | 5.5% | 0.82 s | 0.96 s | 90% | 4.7 GB | Apache-2.0 |
| 20 | Kyutai STT 1BKyutai | 25 | 43.2% | 0.18 s | 0.43 s | 100% | 0.7 GB | MIT |
| 21 | VibeVoice-ASR-Streaming 1.5BMicrosoft | 20 | 39.2% | 0.32 s | 0.51 s | 100% | 1.9 GB | MIT |
English only for now, from 8 speakers (158 spoken lines in all). With a sample this size, treat scores within a few points of each other as ties. Everyone read the same twenty things people say to a voice assistant. The score, out of 100, weighs accuracy most, then speed. Word errors count every word a model got wrong, missed or added. The wait is how long after each line its full text arrived, on Apple M5 Max. Pick a model to see every line it heard.