Voice and speech¶

Use this guide to add a spoken Discord workspace or transcribe incoming audio.

Discord voice workspace¶

Pynchy supports one existing voice channel per Discord connection. Configure the guild and channel by name as shown in Discord, then bind that channel to a normal workspace. Pynchy never creates a voice channel, and it keeps the internal ID in runtime state.

Joining the configured channel activates a voice session. Its user and role allowlists still apply. Pynchy leaves when the last allowed member leaves. Only final agent replies play as speech; streaming text, tool traces, and status messages remain silent.

Install the discord extra, ffmpeg, and a system libopus library. Install and operate Pocket TTS through Local speech synthesis. Pynchy reports its readiness in the speech section of /status.

Pocket TTS listens on 127.0.0.1:8000. Pynchy encodes its WAV output to Discord Opus with FFmpeg. If synthesis is unavailable, it logs the failure and does not fall back to a system voice. Set PYNCHY_DISCORD_OPUS_LIBRARY to an absolute path only when the host installs Opus outside the usual system locations.

DM pairing, interactive question widgets, additional voice rooms, Discord video, and Stage channels are not supported.

Inbound speech-to-text¶

Pynchy can cache supported incoming audio attachments, transcribe them, and replace the message text before the agent sees the turn. Attachment metadata records the cached path and transcription status.

The host first uses the faster-whisper Python package when installed. If it is unavailable, it uses PYNCHY_LOCAL_STT_COMMAND or a whisper executable on PATH. Set PYNCHY_LOCAL_STT_MODEL and PYNCHY_LOCAL_STT_LANGUAGE to tune the local defaults. Discord provides the first built-in integration; other channels need media-download support before they can opt in.

Want to customize this? Write your own channel or speech plugin — see the Plugin Authoring Guide. Have an idea but don't want to build it? Open a feature request.