Local speech recognition
faster-whisper processes Japanese speech on the computer. GPU processing and CPU fallback are supported; actual speed depends on the hardware and selected model.
Software project · Windows · Local AI
Kiwami Reader
Japanese listening practice,
alongside the game.
A Windows desktop companion that captures system or microphone audio and displays Japanese captions with rōmaji in an overlay. The locally documented version 2.2 combines speech recognition with vocabulary tools to support Japanese listening and study.
Live captionsLocal vocabulary lookupSaved wordsOptional English
How it works
Listen. Read. Revisit.
Use Windows output capture for PC games and other audio, or microphone input for a console or TV. Captions appear as drafts and are refined after pauses in speech.
Words in context
Study mode freezes the caption and makes words clickable. Local dictionary lookup provides readings, rōmaji, and possible English meanings. Script colors distinguish kanji, hiragana, and katakana.
Build a vocabulary list
Save words with example sentences and export them as CSV. Correct a final Japanese caption to refresh its rōmaji and glossary within the current session.
Implementation
Built around
local processing.
Python, Qt for Python, faster-whisper, Windows audio capture, and a local Japanese dictionary provide the core desktop workflow.
- No account or API key requiredInitial installation and model downloads need internet access. Speech recognition and dictionary lookup then run locally.
- Optional English translationA separate local Whisper worker translates finalized audio. It adds processing load and may arrive after the Japanese caption.
- Designed for learningCaptions can be delayed or inaccurate, especially with music, names, dialect, or overlapping speech. Dictionary meanings are possible senses, not verified translations of the sentence.
Project status
A desktop prototype.
This page describes the local version 2.2. A public source repository and downloadable release are not linked here yet. No live-audio performance benchmark is claimed.
Raw audio is not saved by the application. Saved vocabulary and preferences persist locally; system audio capture includes other apps using the selected output.