Voice messages

Talk instead of typing - in the browser with the mic button, and from your phone with Telegram voice notes. Claude cannot listen to audio, so your recording is first turned into text on your server; the transcript is what Claude reads.

Choose an engine (Settings -> Voice)

API keys are stored in the Keys vault (you can paste them directly in Settings -> Voice).

Local engine: models and speed

Pick the model in Settings -> Voice, then Install / Download (runs in the background, with progress; you can leave the page). Without a download, the first voice message downloads the model.

Order of magnitude on 2 vCPUs: a 1-minute message takes ~15 s with base, ~30-60 s with small (faster with more cores). The model stays in memory for 10 minutes after the last message, then is unloaded. One message is transcribed at a time; the others wait their turn.

Language: automatic detection works well for normal speech. Fix the language (Settings -> Voice -> Language) if you always speak the same one, or for very short or noisy notes.

In the browser

• The mic button sits next to Send in every session.
• Tap to start, tap ✓ to stop; or hold to talk and release to stop (push-to-talk). Slide left while holding, press Esc or ✕ to cancel.
• When you stop, the voice message is sent right away, like on Telegram: it appears in the conversation as one bubble: a round mic button (tap it to listen to the recording, again to pause) and, next to it, the transcript once the server has made it; the transcript then goes to Claude. You can keep talking or typing meanwhile: the next messages wait their turn. The recording is saved in the topic (attachments/YYYY-MM/voice-….webm).
• Prefer to read the text before sending? Turn off Settings -> Voice -> Send voice messages right away: the transcript is then inserted at the cursor so you can fix it, or sent as text with Send automatically.
• If the transcription fails (or nothing was heard), the recording stays in the conversation and nothing is sent to Claude; send it again or type your message.
• Browsers only give access to the microphone on HTTPS pages (or localhost). With a domain (Caddy) it just works; over plain http://<tailscale-ip> it does not - use an HTTPS access mode or an SSH tunnel to localhost.
• Works on iPhone (Safari) and Android (Chrome). Allow the microphone when the browser asks; if you refused once, re-allow it in the site settings (the lock icon in the address bar).
• Recordings are limited to 15 minutes.
• Record a test in Settings -> Voice shows the transcript, the detected language and the time it took.

On Telegram

Send a voice note, an audio file or a round video note in a session thread: 1. the bot replies with 🎤 <transcript>, 2. the transcript is sent to the topic's session as your message (the audio file is attached, saved under <topic>/attachments/voice/YYYY-MM/), 3. Claude answers as usual.

If voice messages are off, the bot replies with a short hint instead.

From the terminal

maxxing voice status                 # engine, model, readiness
maxxing voice set --engine local --model small --language auto
maxxing voice download [model]       # pre-download a local model
maxxing voice transcribe note.ogg    # test a file

maxxing doctor reports the voice status. To install the local engine by hand: cd /opt/claude-maxxing/server && uv sync --extra voice (then maxxing update keeps it).

Troubleshooting

• Mic button greyed out - voice is off or not ready: open Settings -> Voice.
• "Browsers only allow the microphone on HTTPS pages" - see above.
• "ffmpeg is missing" - apt-get install -y ffmpeg.
• Wrong language / gibberish - fix the language, or use a bigger model.
• Slow - use base, or switch to Groq.