Show HN: Nano-Edge-TTS – 70x realtime, Lightweight Kokoro-based TTS model (huggingface.co)

🤖 AI Summary
A new lightweight text-to-speech (TTS) model known as Nano-Edge-TTS has been introduced, significantly reducing the size and computational demands of the Kokoro-82M model. This innovative decoder, with just 1.75 million parameters, replaces Kokoro's original 53.3 million-parameter waveform decoder with two optimized networks operating at 80 and 160 Hz, enhancing real-time performance by up to 70 times. This allows the model to maintain high audio quality while drastically lowering memory usage and processing time, making it suitable for edge devices. The significance of Nano-Edge-TTS lies in its ability to deliver efficient speech synthesis without sacrificing the quality of prosody and pronunciation, as it retains the unchanged text processing side of the original Kokoro model. Though some voiced sounds may present a slight machine-like tone, it largely performs comparably in auditory tests. This advancement could pave the way for broader applications of TTS technology in less resource-intensive environments, thereby democratizing access to high-quality speech synthesis in mobile and embedded systems. The work is detailed in a paper that discusses the model's architecture, training methodology, and performance evaluations.
Loading comments...
loading comments...