emzod/speak-turbo

Overview

This skill gives your agent ultra-fast, local text-to-speech so it can speak in real time. It delivers ~90ms time-to-first-sound after the daemon is warmed, with eight built-in voices and options to play audio or save to a file. The daemon auto-starts and auto-shuts down after idle to keep resource use practical.

How this skill works

The skill runs a small Rust CLI that streams text to a Python daemon hosting a lightweight TTS model in memory. On first run the daemon loads the model (2–5s), then subsequent requests stream audio and begin playback in ~90ms. It exposes a local HTTP streaming endpoint and supports immediate playback or writing WAV output files.

When to use it

• You need instant audible feedback from an AI agent (~90ms latency).
• You want local-first TTS without sending audio to cloud services.
• You prefer a small, fast TTS engine with several ready-made voices.
• You need streamed playback while long text generates.
• You want a simple CLI/integration for rapid prototyping of voice-enabled agents.

Best practices

• Warm the daemon with an initial call at app startup to avoid the 2–5s first-run delay.
• Use the built-in voices for low-latency responses; switch to a voice-cloning tool if you need custom personalities.
• For reproducible audio, save to WAV (-o) and verify playback when debugging audio issues.
• Suppress status messages in production with quiet mode (-q) while keeping audio playback enabled.
• Monitor the daemon health endpoint (127.0.0.1:7125/health) for readiness and graceful restarts.

Example use cases

• Real-time agent replies in a desktop assistant where sub-100ms responsiveness is required.
• Voice-enabled chatbot in a kiosk or local appliance without cloud audio.
• Rapid prototyping of spoken UX flows using different built-in voices.
• Automated alerts that speak status messages locally, with option to save recordings.
• Offline demos where speed and small footprint matter more than cloned voices.