DRU-651 -- Voice notes reach the chat agent as text

[czpython]

A person speaks a voice note in WhatsApp, or sends an audio clip on Slack, and the agent answers what they said, in text. Nothing shows on the channel before the reply.

• A SpeechToText card in chat/services.py: url, key, and model, not required, no check on save. transcribe posts the audio as multipart to {url}/audio/transcriptions with the model and returns the answer's text. OpenAI, Groq, and a local server that speaks the same shape all work. An empty answer is a TranscriptionError, a new typed error in chat/exceptions.py.
• Message.transcript, one text column, empty by default, with its migration. body stays what the person typed, file the audio.
• In send_turn, before the blocks: for each pending message with an audio/* file, Druks calls the card, saves the transcript, and commits it with the delivered state, before the prompt.
• The message's text block is the body, then a [Voice note] line, then the transcript, empty parts left out. VOICE_NOTE_MARKER lives in chat/constants.py beside the internal message templates. Audio adds no upload and no link.
• When the call fails, no card is connected, the answer is empty, or the file is over MAX_UPLOAD_BYTES, the transcript stays empty and Druks adds an internal message from TRANSCRIPTION_FAILED_MESSAGE to the same turn, worded neutrally. The notes close the turn. The turn runs. A note over the cap makes no call.
• The message response and the web page show the transcript under the typed text.
• docs/configuration.md describes the Speech To Text card. docs/chat.md describes voice notes under WhatsApp and Slack, and what happens when Druks cannot transcribe.

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated
druks	🟢 Ready	View Preview	Sep 26, 2026, 8:26 PM

💡 Tip: Enable Automations to automatically generate PRs for you.