Piper’s native speech library starts returning audio before it has synthesized an entire sentence. Applications such as screen readers can begin speaking sooner when they adopt the updated library and prepare a voice for streaming.
In a test using a rebuilt Piper macOS app on an Apple M3 running macOS 26, the first audio buffer for a six-second sentence arrived after 0.24 seconds instead of 4.15. The measurement used the speech-extension path used by VoiceOver and wrote audio to files; it measured buffer delivery, rather than sound reaching a listener.
The change also lets the library use several processor cores. Streaming requires a voice split into two model files with the supplied tool; ordinary voices keep sentence-at-a-time output. Contributor herwigfelix makes a first contribution to Piper with the work.