kesha-voice-kit, created by Drakulavich, gives LLM-based agents speech input and output. The tool supplies local speech-to-text and text-to-speech processing for agent interaction and desktop assistants. Key elements include 25-language STT, Kokoro and Piper TTS engines, SSML prosody controls, and a compact ~65MB Rust engine. It targets developers, Claude Desktop and OpenClaw power users, and privacy-conscious individuals needing offline voice in agent workflows.

What tasks you can actually use it for?

The kit converts spoken audio into text and generates spoken audio from text, so it fills the gap between text-only LLMs and voice-enabled agents. It includes voice activity detection and automatic language detection for live input, making it suitable for live assistant interactions, voice commands, and dialog testing with agent frameworks. The toolkit aims to let an LLM both hear and speak within an agent workflow.

How fast and lightweight the processing feels in practice

Speech-to-text is optimized for Apple Silicon, with the developer reporting up to 19x faster performance than a standard Whisper implementation, and the STT engine supports 25 languages. The core engine is a single approximately 65 MB binary that runs locally, which reduces startup footprint and removes heavyweight runtime dependencies normally associated with media stacks.

What voice control and language options are available

Text-to-speech uses Kokoro and Piper engines to produce natural-sounding output and supports SSML to adjust rate, pitch, and volume. TTS is available in nine languages while STT covers 25 languages, which lets developers choose voice parameters programmatically for dialogue testing, prototypes, and multilingual agent interactions.

Does it fit into an agent workflow without extra dependencies?

The kit functions as an MCP server and provides direct integration points for Claude Desktop, OpenClaw skills, and Hermes agent protocols. It is compatible with Apple Silicon via CoreML and with Linux and Windows via ONNX. CLI tools require Bunor newer, and the engine avoids Python, FFmpeg, and native Node addons by shipping as a Rust binary. Initial model downloads are needed during setup; after that processing stays local.

Best suited for developers needing local, agent-integrated voice

The kit is a practical option for developers and power users who require offline voice in agent environments, supported by active development from the creator and strong community adoption among OpenClaw and Claude Desktop users. Expect a modest setup step for models and Bun, and plan to verify voice behavior in your specific agent flows before deploying to production environments.