other / agent
backtalk
A voice interface for Claude Code: hold a key, speak and hear the reply read aloud, with speech recognition and a built-in voice that run locally.
At a glance
- Purpose
- A voice interface for Claude Code: hold a key, speak and hear the reply read aloud, with speech recognition and a built-in voice that run locally.
- Runs in
- Local agent with shell
- Operating systems
- macOS, Linux, Windows
- Needs
- Runtime: Python environment, uv and espeak-ng; Needs internet: No; Needs accounts or API keys beyond your agent: Not stated
- Files and data
- Reads files: Yes; Writes files: Yes; Sends data to other services: Yes
- Tool type
- Agent
- Delivery
- From GitHub
- Version
- Latest on the main branch
- Access
- Code on GitHub
- Compatibility
- According to the README, backtalk runs on Claude Code only and is built on the Claude Agent SDK. install.sh and run.sh are for Mac and Linux, and Windows setup runs through a wizard in Claude Code. Music ducking is described for Spotify on macOS. For the optional ElevenLabs voice, the README describes the macOS Keychain, secret-tool on Linux and an environment variable on Windows for the key. Speech recognition uses faster-whisper, and Kokoro provides the built-in voice.
- External costs
- The README requires Claude Code with a Claude subscription and says spoken turns use that plan's usage. The README describes local speech recognition and Kokoro voice models without voice API keys or per-word charges. The README describes ElevenLabs as an alternative using your account and API key, with a free audition tier and a paid plan for day-to-day use. Your coding agent or model provider may charge for usage.
- Permissions
- The README says installation creates a Python environment, downloads speech/voice models and installs a system library; uv installation is offered. The README says the Windows wizard installs uv, espeak-ng, environment and models. The README requires authenticated Claude Code subscription access. The README describes microphone input and a system-wide keyboard listener requiring macOS Input Monitoring. The README says push-to-talk defaults on; hands-free keeps voice detection listening. The README describes agent file changes, including vault notes, with spoken/typed approval prompts and denial on timeout; bypassPermissions lets the agent act without asking. The README says voice commands save microphone/approval settings. The README describes ElevenLabs keys via macOS Keychain, Linux secret-tool or environment, requiring ffmpeg for that voice. The README says backtalk writes tiny state files while it listens, thinks and speaks.
- Usage terms
- GNU Affero General Public License v3.0, as stated in the repository. See its LICENSE file for the exact terms. The README specifies version 3 or later, permits commercial use, and describes source-sharing obligations for redistribution and modified services used by others. The README asks for author credit and directs closed-source commercial-product inquiries to the author. The README identifies faster-whisper and Whisper models as MIT, Kokoro as Apache 2.0, and espeak-ng as GPL-3.0 used as a system tool.
Code on GitHub
jaredrhod/backtalk
AGPL-3.0Talk to your Claude Code agent out loud. Hold a key, speak, and it answers in a real voice about a second later, with all its tools and its own personality. The hearing and the voice run on free local models, no API keys.
Topics: claude claude-code kokoro push-to-talk python speech-recognition text-to-speech voice
Public GitHub data fetched · 0 days old
View on GitHub ↗What it does
backtalk is a voice interface for Claude Code. The README describes holding a key to speak, local transcription into a live Claude Code session, and sentence-by-sentence spoken replies.The README says the session uses the selected agent folder and its CLAUDE.md. The README describes spoken approval prompts, yes to approve, details to hear the command, and other responses to deny or steer it. The README describes voice commands for session clearing, compaction, model switching, microphone modes and approval settings.
The README lists faster-whisper speech recognition, a local Kokoro voice and an ElevenLabs alternative using your API key. The README says state files expose listening, thinking and speaking status to other programs.
Who it is for
People who already use Claude Code with a Claude subscription and want to speak to their agent and hear its replies.Example tasks
Hold the talk key, ask your Claude Code agent a question and listen to the spoken reply.Answer a spoken permission request with yes, ask for details, or say no with a reason.
Say go hands free to switch to hands-free listening.
Set resume_last_session to true in the config to reattach to your previous conversation at launch.
Point agent_dir at the folder whose CLAUDE.md is your agent.
Getting started
The README says to clone the repository, change into the backtalk folder and run ./install.sh. The installer sets up a Python environment, two local models for speech recognition and the voice, and one system library. To configure it, open the folder in Claude Code and say 'read backtalk.md and set me up.' Or copy backtalk.json.example to backtalk.json, edit it, and run ./run.sh. On Windows the README says to use the wizard instead of the scripts. The README has the exact commands.Limitations
According to the README, backtalk runs on Claude Code only and needs Claude Code with a Claude subscription, plus uv, which the installer offers to install. The README says every spoken turn uses your Claude Code plan and the first run downloads about 1 GB of models. The README warns that connecting a headset can switch the default microphone to a narrowband call profile and degrade audio; it describes setting mic_device to select a microphone. The install.sh and run.sh scripts are for Mac and Linux, and Windows setup runs through a wizard in Claude Code instead. The talk key needs a global key listener, and the README says macOS asks for Input Monitoring permission. The optional ElevenLabs voice needs your own account, an API key and ffmpeg. Music ducking is described for Spotify on macOS. The repository can change at any time; check its README before you install.Data practices
The README says speech recognition and the built-in Kokoro voice run locally, while transcribed words enter Claude Code and use its session. The README describes ElevenLabs as a cloud voice alternative, with Kokoro fallback and failures recorded in logs/backtalk.log. Using that voice requires sending material for speech synthesis to ElevenLabs; the README does not detail the request payload.The README describes key storage through macOS Keychain, Linux secret-tool or the ELEVENLABS_API_KEY environment variable, including that variable on Windows. The README says configuration, keys and agent identity are outside tracked files and describes listening, thinking and speaking state files for other programs. The README does not describe telemetry.