smallest-voice-bridge

Discord voice bridge extension for OpenClaw with:

• wakeword detection via OpenWakeWord
• one-shot buffered STT via Smallest.ai
• TTS replies in Discord voice
• recording mode for event / meeting capture

This extension is designed for voice-first OpenClaw workflows where you want to join a Discord voice channel, wake the assistant with a keyword, speak naturally, and get a spoken reply.

Features

• /join command to bring the bot into your current Discord voice channel
• Wakeword-triggered voice prompts
• Buffered single-shot prompt transcription after silence
• Spoken responses via Smallest.ai TTS
• Recording mode with transcript and summary pipeline
• Status command handling during recording mode

Requirements

• OpenClaw >=2026.3.22
• Node.js compatible with your OpenClaw install
• Python 3.10+
• ffmpeg available on PATH
• Smallest.ai API key
• Discord bot token configured in OpenClaw

Python dependencies

Install the wakeword-side dependencies in your Python environment:

pip install openwakeword numpy onnxruntime

Depending on your system, you may also need:

pip install scipy

Install

Option 1: local development install

Clone the repo into your OpenClaw extensions folder:

git clone <YOUR_GITHUB_REPO_URL> "%USERPROFILE%\\.openclaw\\extensions\\smallest-voice-bridge"

Option 2: OpenClaw extension path install

If your OpenClaw setup supports extension path installs, point it at this folder/repo and enable the plugin in config.

OpenClaw config

Example plugin config:

{
  "plugins": {
    "entries": {
      "smallest-voice-bridge": {
        "enabled": true,
        "config": {
          "apiKey": "YOUR_SMALLEST_API_KEY",
          "wakeword": {
            "enabled": true,
            "pythonBin": "python",
            "scriptPath": "C:\\Users\\<you>\\.openclaw\\extensions\\smallest-voice-bridge\\oww_stream.py",
            "modelPath": "",
            "modelName": "hey mycroft",
            "threshold": 0.6,
            "vadThreshold": 0.5,
            "patience": 3,
            "debounceSec": 1.5,
            "commandWindowMs": 6000,
            "minAutoReplyScore": 0.85,
            "wakeCooldownMs": 2500,
            "ttsSuppressWakeMs": 1200,
            "minPromptChars": 6,
            "shortPromptHoldMs": 1200,
            "repromptText": "",
            "ackMode": "earcon"
          }
        }
      }
    }
  }
}

Discord config

Make sure OpenClaw Discord is enabled and your bot token is set.

Usage

1. In Discord text chat, run:

/join

2. Join the same voice channel as the bot.
3. Say the configured wakeword, for example:

hey mycroft

4. After the earcon, say your prompt.

Recording mode

To start recording:

start recording

To stop recording:

stop recording

To ask for progress / status:

status

Recording artifacts are stored under:

~/.openclaw/workspace/notes/event-recordings/YYYY-MM-DD/

Notes

• Wakeword behavior depends a lot on mic quality, room acoustics, and pronunciation.
• ffmpeg must be installed and accessible from the shell OpenClaw uses.
• This extension currently assumes mono 16 kHz PCM for wakeword and prompt capture.
• During recording mode, conversation should be treated as transcript-only and should not route as a live assistant prompt.

Troubleshooting

Wakeword sidecar spawned but no detection

• verify Python is available
• verify openwakeword imports correctly
• verify model name / model path

STT fails

• verify Smallest API key
• verify outbound network access
• verify ffmpeg is installed

Bot joins but no voice response

• verify Discord voice permissions
• verify TTS connectivity
• check OpenClaw logs for player state transitions

Development

Main files:

• index.js — OpenClaw extension entry point
• src/voice-session.js — wake -> capture -> STT -> routing flow
• src/wakeword.js — wakeword sidecar management and gating
• src/stt.js — full-audio transcription helper
• src/tts.js — speech synthesis and playback
• src/recording.js — recording mode and artifact generation
• oww_stream.py — OpenWakeWord streaming sidecar

License

MIT