Paseo Korean Voice

English · 한국어

A Paseo plugin that makes dictation and voice mode work properly in Korean. It runs a local OpenAI-compatible speech gateway: transcriptions go to Groq/OpenAI Whisper with a Korean prompt and glossary, and replies are spoken with macOS Korean voices.

Why Korean breaks (Paseo 0.11.2)

• Local speech models cover English (parakeet v2, kokoro-en) and 25 European languages (parakeet v3). There is no local Korean STT or TTS model.
• Switching STT to an OpenAI-compatible provider (Groq Whisper, etc.) with language: "ko" makes dictation work. However, the daemon always sends Whisper a fixed English prompt ("Transcribe only what the speaker says…"), so you can't supply Korean context or a glossary.
• Whisper tends to invent Korean subtitle boilerplate such as "시청해 주셔서 감사합니다" ("Thanks for watching") on silence.
• Voice mode TTS defaults to kokoro-en, which can't read Korean replies. OpenAI TTS works but needs a paid key.

How it works

The plugin's server process serves OpenAI-compatible endpoints on 127.0.0.1:6791, and Paseo's OpenAI speech settings point at it.

Paseo daemon ──/v1/audio/transcriptions──▶ korean-voice ──(Korean prompt + glossary)──▶ Groq/OpenAI Whisper
             ◀────── text with hallucinated sentences removed ──────┘
Paseo daemon ──/v1/audio/speech (pcm)──▶ korean-voice ──▶ macOS `say` (Yuna and other Korean voices, free and offline)

• STT requests: replaces Paseo's English prompt with a Korean prompt and glossary, and adds language=ko if missing. API keys are never stored; the Authorization header Paseo sends is forwarded upstream as-is.
• STT responses: drops known Whisper hallucinations sentence by sentence.
• TTS: strips Markdown, code blocks, and URLs, then renders 24 kHz PCM with a macOS Korean voice.
• Settings screen: Settings → Plugins → Korean Voice changes voice, rate, glossary, and upstream URL, and plays a test phrase.

The prompt and glossary are plain settings, so the same approach works for Japanese, Chinese, or any other language Whisper supports.

Install

Requirements: macOS, Paseo 0.11.2+, and a Groq or OpenAI API key.

# 1) Turn on Settings → Plugins → "Enable plugins"
#    (or add "pluginsEnabled": true to ~/.paseo/config.json and run `paseo reload`)

# 2) Install the plugin
paseo plugin install npm:paseo-korean-voice      # or github:me0w2en/paseo-korean-voice
paseo plugin ls            # korean-voice should be running

# 3) Point Paseo speech at the gateway (backs up config.json, then merges), then restart the Paseo app.
#    providers.openai.stt.apiKey in ~/.paseo/config.json must hold your Groq/OpenAI key.
git clone https://github.com/me0w2en/paseo-korean-voice.git
node paseo-korean-voice/scripts/apply-voice-config.mjs --dry-run   # preview
node paseo-korean-voice/scripts/apply-voice-config.mjs

Merge this into ~/.paseo/config.json, then restart the daemon:

{
  "features": {
    "dictation": { "stt": { "provider": "openai", "language": "ko" } },
    "voiceMode": {
      "stt": { "provider": "openai", "language": "ko" },
      "tts": { "provider": "openai" }
    }
  },
  "providers": {
    "openai": {
      "stt": { "baseUrl": "http://127.0.0.1:6791/v1", "apiKey": "<your Groq or OpenAI key>" },
      "tts": { "baseUrl": "http://127.0.0.1:6791/v1", "apiKey": "local" }
    }
  }
}

The gateway forwards to https://api.groq.com/openai/v1 by default; change Upstream STT URL in the plugin settings for OpenAI or a local Whisper server.

To develop, install from a local clone (paseo plugin install "$PWD"), run npm install && npm run typecheck, and apply changes with paseo plugin reload korean-voice.

Uninstall

• Restore ~/.paseo/config.json.bak-korean-voice-* and restart the daemon, then paseo plugin remove korean-voice.
• If you disable the plugin while step 3's settings are still in place, dictation fails. Revert the speech settings together with the plugin.

Limitations

• macOS only: TTS uses say. If your daemon runs on a remote Linux host, TTS won't work.
• STT still uses cloud Whisper (Groq/OpenAI). For fully offline use, point the upstream URL at a local OpenAI-compatible server such as whisper.cpp.
• Voice mode's system instructions are hard-coded in English. Agents usually answer in Korean when you speak Korean.

License

MIT