Vincent Voice Chat

A Whisper-, Kokoro- and Opencode-powered voice chat Agent CLI.

Voice chat with your AI agent:

• records microphone audio until you press Enter,
• transcribes with local Whisper,
• sends each turn to opencode run,
• speaks assistant replies with Kokoro.

Setup

uv venv --python 3.13 --seed
uv sync

Use Python 3.13 for best Kokoro compatibility.

Troubleshooting

• No module named pip

  • Recreate the venv with seeding: uv venv --python 3.13 --seed && uv sync
  • Quick repair in existing env: uv run python -m ensurepip --upgrade

• Warning: You are sending unauthenticated requests to the HF Hub

  • Optional only. Set HF_TOKEN if you want better download limits: export HF_TOKEN=hf_your_token_here

• No audio output from Kokoro

  • Check detected devices: uv run python -c "import sounddevice as sd; print(sd.default.device); print(sd.query_devices())"
  • Verify with text-only mode to isolate TTS/device issues: uv run vincent --no-voice

Usage

uv run vincent
uv run vincent --no-voice  # Disable assistant speech output

# Useful options
uv run vincent --help
uv run vincent --new-session
uv run vincent --session-id ses_abc123
uv run vincent --whisper-task translate
uv run vincent --whisper-model small
uv run vincent --input-language en
uv run vincent --input-sample-rate 16000 --input-channels 1
uv run vincent --keep-input-audio
uv run vincent --tts-voice af_heart --tts-lang-code a --tts-speed 1.0

# Look up Kokoro Language Codes, Voices and Whatnot
uv run kokoro-info --lang-codes
uv run kokoro-info --aliases
uv run kokoro-info --voices

Say exit or quit to stop.

Input Flags

• --input-language: hint the spoken language for Whisper (faster and often more accurate); omit to auto-detect.
• --input-sample-rate: microphone capture rate in Hz (default 16000, good for speech).
• --input-channels: microphone channel count (1 mono is typical; 2 stereo if needed).
• --keep-input-audio: keep each turn's WAV in .voice_inputs/<session>/.