WhisperPress — offline push-to-talk dictation app for Windows built on whisper.cpp

[b84330808]

First of all — thank you @ggerganov and all contributors for this amazing project. whisper.cpp is the entire engine room of what I'm sharing here.

WhisperPress is a free, MIT-licensed dictation app for Windows: hold a hotkey, speak, release — the transcript is typed into whatever app your cursor is in. Everything runs locally.

_2026_06_12_02_21_16_302.mp4

Repo: https://github.com/b84330808/whisperpress

How whisper.cpp is used:

• The app downloads the official release binaries (v1.8.6, BLAS or cuBLAS flavor) on first run — no bundled binaries, no compilation, and users pick a GGML model which is fetched from the HF repo.
• whisper-server stays resident with the model loaded, so dictation latency is just inference (sub-second for short utterances on CUDA with large-v3-turbo-q5_0).
• Long audio files / meeting recordings go through whisper-cli with -oj -pp -np, parsing the segment lines from stdout to stream live results into the UI while it transcribes.

Two findings that may be useful to other whisper.cpp users:

1. CJK punctuation: whisper-server defaults to greedy decoding (-bs -1), unlike whisper-cli (beam 5). With greedy decoding,
   Chinese output very often loses all punctuation. Passing -bs 5 -bo 5 to the server brought it back. A punctuated, conversational initial prompt helps further — prompts are style mimicry, so an unpunctuated prompt (or an "instruction" like "add punctuation") actually makes things worse.
2. Readiness probe: the server loads the model before binding the port, so a simple GET / returning 200 is a reliable "model is ready" signal — handy for supervising it as a child process.

Other bits: meeting transcription via WASAPI loopback, deterministic Traditional/Simplified Chinese with OpenCC on top of whisper output, and SRT export from the segment timestamps.

Feedback is very welcome — especially from anyone dictating in CJK languages. And thanks again for whisper.cpp!