A hands-free voice loop for Claude:
mic → ElevenLabs Scribe (STT) → Claude (local `claude` CLI) → ElevenLabs (TTS) → speaker
Why not Whisper / "use Claude to transcribe"? Claude has no audio input — it can't transcribe. So a fast hosted speech model (ElevenLabs Scribe) does the STT, and Claude is the brain in the middle. This avoids running Whisper locally (slower, heavier).
Auth (no new credentials)
Reuses the ElevenLabs API key already set for eleven-labs-skill
(~/.claude/skills/eleven-labs-skill/config.json, or ELEVENLABS_API_KEY).
The brain uses your existing Claude Code login via the claude CLI — no API key needed.
One-time setup
pip3 install -r ~/.claude/skills/voice-mode/requirements.txt # sounddevice, numpy, requests
python3 ~/.claude/skills/voice-mode/voice_mode.py setup # verifies ElevenLabs + claude CLI
macOS will prompt for microphone permission the first time the terminal records.
The sounddevice wheel bundles PortAudio — no brew install needed.
Use
# Continuous conversation (default). Speak, pause, hear the reply. Say "goodbye" or Ctrl-C to stop.
python3 ~/.claude/skills/voice-mode/voice_mode.py converse
# One turn: record → answer → speak
python3 ~/.claude/skills/voice-mode/voice_mode.py once
# Utilities
python3 ~/.claude/skills/voice-mode/voice_mode.py listen # record → {"text": ...}
python3 ~/.claude/skills/voice-mode/voice_mode.py say "hello" # TTS only
python3 ~/.claude/skills/voice-mode/voice_mode.py stt clip.wav # transcribe a file
python3 ~/.claude/skills/voice-mode/voice_mode.py mics # list input devices
It auto-detects when you stop talking (silence-based cutoff), so there's no push-to-talk.
Config (config.json)
| key | default | notes |
|---|---|---|
voice_id |
21m00Tcm4TlvDq8ikWAM |
ElevenLabs voice (Rachel, a standard voice) |
stt_model |
scribe_v1 |
ElevenLabs Scribe |
tts_model |
eleven_turbo_v2_5 |
low-latency; use eleven_flash_v2_5 for min latency |
brain |
claude |
claude (Claude Code CLI) or codex |
silence_secs |
1.2 |
trailing silence that ends an utterance |
max_utterance_secs |
30 |
hard cap per turn |
Change the voice: voice_mode.py setup --voice-id <ID> --voice-name "<name>"
(browse voices with the eleven-labs-skill). Idan already has a personal voice clone
from the digital-twin skill — pass its voice id here to have Claude reply in his voice.
Switch the brain to Codex: voice_mode.py setup --brain codex.
Notes / gotchas
- Replies are constrained by a system prompt to be short and TTS-friendly (no markdown/lists).
- Each turn sends the running transcript to
claude -p(fresh context per call — avoids the--continuestaleness issue), keeping the last ~6 exchanges. - STT and TTS are hosted calls (billed to your ElevenLabs account); latency is network-bound, typically a second or two per leg — far faster than local Whisper on this machine.