Live conversation can be a barrier for many people in our communities. When speaking is difficult, or someone needs more time to put their thoughts into words, a meeting can move on before they have a chance to contribute. People need a way to take part, express themselves in their own voice and stay in control of what they say.
I built a Mac app that lets users type and speak in their own cloned voice during meetings, with control over exactly when an answer is spoken. Users connect their ElevenLabs voice, and the app routes speech through Core Audio and BlackHole into the meeting microphone.
Sentence completion and answer drafting run locally on their Mac. Users can also save an approved audio take with an answer and reuse that exact recording.
The architecture, at a glance.
Composer, local completion and saved answers
ElevenLabs voice or an approved saved take
PCM audio checks and a dedicated output path
BlackHole routes speech into the call
Qwen3-TTS and VoxCPM2 can also generate speech locally through the same audio path when their model files are installed.
A live conversation moves quickly, so finding the words and choosing when to speak both matter. Sentence completion and saved answers help users prepare a response. They can press Speak without waiting for completion, and drafting an answer never speaks it automatically. Escape stops queued and active speech, while approved audio takes let users reuse phrases they’ve already checked.
If speech generation or the audio device fails during a meeting, users need to recover without losing their answer or saying it twice. The app checks for silence and unusable audio, with timeouts and cancellation if something gets stuck. Failed text stays available for the user to retry. If playback has already started, the app won’t automatically repeat the answer.
I added timing measurements and benchmark scripts for the local speech engines, including waveform comparisons when testing inference optimizations. Those checks are separate from listening to voice quality and testing what other participants actually hear in a call.