Speed and accuracy columns are relative ratings from 1 to 10, for comparing models at a glance. Measured word error rates and latency come from our benchmark suite: see How We Benchmark for the method, or the leaderboard for every number.
Superwhisper (Cloud)
Hosted by Superwhisper and optimized for low latency and accuracy. Works in 100+ languages with nothing to download.
S1-Voice has the lowest word error rate on our benchmarks, 6.6% across the eight benchmark datasets. See How We Benchmark for what that measures.
Cohere Transcribe (Local)
The most accurate local model, with an 8.4% word error rate on Apple M4. It supports 14 languages (listed above), doesn’t translate, and doesn’t support realtime transcription. It runs through MLX at 4-bit precision on your device’s GPU: Apple silicon on macOS, Vulkan on Windows. It isn’t available on Windows ARM.Nvidia Parakeet (Local)
Based on Nvidia’s Parakeet models, running locally. On macOS they run through Argmax’s WhisperKit SDK; on Windows they run through ONNX. They process long recordings in parallel. They can struggle with punctuation and occasionally hallucinate on single-word recordings.
Parakeet Multilanguage covers 24 languages. On iOS, a Parakeet Japanese model is also available.
Parakeet Multilanguage is the only local option on Windows ARM (Snapdragon) devices, which don’t support Vulkan. See Windows requirements.
Whisper Models (Local)
Based on the Whisper model series from OpenAI, running locally through whisper.cpp.
Every Whisper model is free, including Large v3 Turbo. They’re the models to reach for on the Free plan, and they run entirely on your device. Whisper Large v3 and the Chinese Turbo build are marked experimental in the app.
These models were renamed in the app. Whisper Large was formerly Superwhisper Ultra, Whisper Large v3 Turbo was Ultra V3 Turbo, and the small models were Standard, Nano, and Fast.