Show HN: Phonon TTS on WebGPU, 44MB, 115x real time, with live word timings (github.com)

The "blixt/phonon" project has unveiled a high-performance text-to-speech (TTS) model capable of generating speech at 115 times real-time speed in web browsers, leveraging WebGPU technology. This model, based on the Phonon 31M architecture by Gradium, is optimized for lightweight deployment—only requiring a 44MB download for the model weights, thus enabling seamless integration into web applications. It offers live timing information for every word and token spoken, enhancing accessibility features by providing precise start and end times for words during playback. This development is significant for the AI/ML community as it showcases the potential of WebGPU in accelerating machine learning models directly within browsers, moving beyond traditional CPU-bound executions. With impressive benchmark results, such as achieving 3ms for the first audio output on an M3 Max, it paves the way for real-time applications in various domains, including education and accessibility technologies. Furthermore, the capability to stream audio while processing ensures a smooth user experience, while the model's ability to adaptively maintain high performance across different device architectures will likely encourage further innovations in TTS and browser-based AI applications.