Speech-to-text is only useful if it feels instant. Pithflow returns cleaned text in under a second for short clips (one sentence) and 1-3 seconds for longer paragraphs. The wait between thinking and seeing your words shorter than a blink.
Most dictation tools are accurate enough. What kills the user experience is the wait. If you have to sit there 5 seconds after releasing the hotkey before text appears, you stop trusting the tool — you start typing instead. Sub-second response keeps voice in the "feels instant" zone where you keep using it.
Two things: (1) Specialized inference hardware — fast transcription runs on accelerator hardware that's purpose-built for this workload. (2) Cleanup happens in parallel — by the time the transcription finishes, the cleanup pass is already running so the user only waits for the slower of the two, not both sequentially.
Most older dictation tools take 5-10 seconds for a sentence. Wispr Flow and Pithflow both run on the same generation of fast inference hardware — they're competitive on latency. The bottleneck isn't the model anymore, it's the network round-trip.
Latency includes your network. On a coffee-shop WiFi or a cellular connection on a bad day, the round-trip to our servers adds 200-500ms. We can't fix physics — but the model itself only adds 100-200ms to the user-perceived latency. Most of what you wait for is the network.