Streaming audio and clips
What you can do.
On Mac, clean up your mic and let background music dip when you speak. Save the last 15 or 30 seconds as a clip.
- Noise reduction and voice filters
- Music that dips when you speak
- 15- and 30-second clips
Installation and release
The Mac version will be released after Spoon approval. You can get release news by email.
macOS 26+ · Apple Silicon. The Windows version is coming soon.
What’s been built
- Mac microphone and virtual cable connectionChecked
- Music that lowers itself when you speakReady
- Video clips with captionsChecked
Technology & architectureFrom microphone to broadcast to clip.Explore the architecture
How it works, in three steps
Noise reduction, broadcast mixing and captioned clips follow one connected audio path.
Tract processes 480-frame model hops before dereverberation and voice filters. A dedicated worker keeps neural computation separate from device callbacks.
Microphone and AI-speech RMS levels drive smooth music ducking. A final master limiter feeds BlackHole or the selected virtual audio cable.
Local Whisper and Silero analyze speech and voice regions. A native encoder or a tested FFmpeg path renders the clip, with progress tied to completed work.
Tract processes 480-frame model hops before dereverberation and voice filters. A dedicated worker keeps neural computation separate from device callbacks.
Security & design principles
Keep callbacks lightweight
Bounded SPSC · Preallocated buffersDevice callbacks use preallocated memory and atomic state. Bounded queues connect them to the separate neural worker.
Local audio, separate GPU speech
Rust / Tract CPU · CUDA / TensorRTMicrophone, music and clips run locally. Optional CosyVoice3 synthesis uses an authenticated GPU server, with recorded responses from Azure T4.
Hardware-specific paths
Apple Silicon · macOS 26+Mac virtual-cable output and captioned clips have been checked. Windows and RTX hardware, encoder compatibility and total latency targets remain under validation.