Voice Keyboard

Self-hosted voice input for Japanese — speak, and get send-ready text.

Audio is transcribed (Whisper / OpenAI / Groq) and optionally refined by an LLM (Claude) into tidy, ready-to-send Japanese — fillers removed, punctuation fixed, proper nouns normalized — while preserving your original tone (です/ます ↔ だ/である ↔ casual). Everything runs on hardware you control; no third-party dictation service ever sees your text.

---

Why

Most speech-to-text tools mangle Japanese: they flatten honorifics, butcher proper nouns, and leave you re-editing every dictation by hand. Voice Keyboard puts a thin refine proxy between the STT engine and your text field. The proxy keeps your tone intact, fixes homophones in context, normalizes 旧字体 → 新字体, and can inject a personal proper-noun dictionary so names and jargon come out right. The result is text you can actually send.

Download

Grab the Android APK from the latest release (voice-keyboard-v0.1.0.apk), or build it from source (see below). You'll also need to run the proxy.

Architecture

flowchart LR
    A[🎙 Android IME / Windows hotkey] -->|audio| B[Refine Proxy<br/>FastAPI]
    B -->|STT| C[Whisper / OpenAI / Groq]
    C -->|raw text| B
    B -->|optional refine| D[Claude]
    D -->|clean text| B
    B -->|send-ready text| A

Loading

The proxy is the brain: clients only record audio and paste the returned text. That keeps refine logic, prompts, and the personal dictionary in one place, so a client never needs to change when the pipeline improves.

Components

Dir	What it is
proxy/	FastAPI server (original work). Receives audio, runs STT, optional LLM refine / conversation mode, returns plain text. All keys via env vars.
whisper-to-input/	Android keyboard (IME). A GPL-3.0 fork of j3soon/whisper-to-input, extended with a custom keypad, a one-tap refine/style button, and a conversation mode.
pc/	Windows client. AutoHotkey + PowerShell — a hotkey to record, send to the proxy, and paste the refined text.

Features

• Tone-preserving refine — casual stays casual, polite stays polite; never silently rewrites your register.
• Style modifiers — one tap to force polite / casual / bullet / summary output.
• Conversation mode — talk to an assistant instead of dictating; optional TTS reply.
• Multi-backend STT — OpenAI gpt-4o-transcribe by default, Groq whisper-large-v3-turbo fallback.
• Personal proper-noun dictionary — inject names, places, and jargon so STT/refine gets them right (kept local, never committed).
• Privacy-first — self-hosted proxy; your audio and text stay on your infrastructure.

Proxy quickstart

cd proxy
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

PROXY_SHARED_SECRET=choose-a-secret \
OPENAI_API_KEY=sk-... \
ANTHROPIC_API_KEY=sk-ant-...  \
uvicorn main:app --host 0.0.0.0 --port 9090

Endpoints

Method	Path	Purpose
POST	/v1/audio/transcriptions	OpenAI-API-shaped STT + refine (bearer auth)
POST	/asr	Whisper-ASR-Webservice-shaped STT + refine
POST	/feedback	Post-edit feedback (optional prompt-improvement loop)
POST	/kanji	Kana → kanji conversion candidates
GET	/health	Backend availability

Configuration

Var	Purpose
PROXY_SHARED_SECRET	required — bearer shared between clients and proxy
OPENAI_API_KEY	OpenAI STT + TTS
GROQ_API_KEY	Groq STT (fallback)
ANTHROPIC_API_KEY	Claude refine / conversation (optional)
STT_PROVIDER	openai (default) or groq
BIND_HOST / BIND_PORT	default 0.0.0.0:9090

Personal dictionary

The refiner can read a proper-noun hint dictionary at dictionary/proper_nouns.json to improve name/term recognition. That directory is git-ignored — it holds personal names and terms, so each user supplies their own. The prompts fall back gracefully when it is absent. See dictionary/ notes in the proxy for the expected shape.

Android: build & install

The IME is a standard Gradle Android project under whisper-to-input/android/. On ARM64 build hosts (e.g. Raspberry Pi), set a local aapt2 override in an un-committed gradle.properties (see the comment in that file).

Important

In the Android app settings, point the endpoint at your proxy and set Postprocessing to No Conversion. Some upstream whisper-to-input builds default this to Convert to Traditional Chinese, which turns Japanese output into garbled Traditional-Chinese text.

Attribution & license

• whisper-to-input/ is a fork of j3soon/whisper-to-input, licensed under GPL-3.0.
• The proxy and Windows client are original work.
• This repository as a whole is released under GPL-3.0 — see LICENSE.