I didn't want to pay a monthly subscription for dictation, so I built one that runs on my GPU
Описание и применение
Создана офлайн‑приложение Wisperno для диктовки на ПК с NVIDIA GPU, без подписки и сетевых запросов. Использует Whisper large‑v3‑turbo и квантизированную языковую модель, обеспечивает 0.82‑1.48 с задержки для 100 слов и 13.5× скорость реального времени в режиме живой транскрипции.
Задача
Need offline, subscription‑free dictation on a PC
Результат
Runs fully offline on NVIDIA GPU, 0.82‑1.48 s latency for 100 words, 97.6‑98.8 % fidelity, 13.5× faster than real time in live mode
Рабочий процесс
- Hold hotkey
- Speak
- Release
- Text pasted into focused app
- Optional polish via local model
Доказательства
It runs fully offline. Whisper large‑v3‑turbo transcribes on the GPU, and a small quantized language model cleans up the text.
A long pause (1.3 seconds or more) starts a new paragraph. I first tried segment‑level timestamps from Whisper to detect the pause, and they could not see it at all. Word‑level timestamps can.
The smaller presets are faster and use less memory. The bigger ones keep a little more fidelity: 97.6% for the two smallest, 98.8% for the two 3B models, measured on raw model output.
Live Transcription is a separate mode for meetings and lectures. My first design re‑decoded only a sliding window of the latest audio. The test suite showed it silently dropped real words. The current version re‑decodes the whole unconfirmed audio buffer every 0.6 seconds. It runs about 13.5 times faster than real time (RTF 0.074), with no second model loaded.
- Дата записи: 2026-10-21T14:41:05.380Z