Nano-Edge-TTS replaces Kokoro-82M’s 53.3-million-parameter waveform decoder with a 1.75-million-parameter harmonic-plus-noise decoder, while keeping Kokoro’s text processing, pronunciation, timing, pitch, and voices unchanged. Rather than running a neural network at the audio sample rate, two small networks operate at 80 and 160 frames per second to control the gains and phase of a pitched pulse train and white noise; inverse STFTs convert these signals into speech. The model is intended as a lightweight, fast alternative that can be loaded into Kokoro’s existing pipeline.
On an Apple M4 CPU, the decoder runs 33 times faster than Kokoro’s, and the full phoneme-to-audio pipeline runs 11 times faster; the whole model uses 30.2 million parameters instead of 81.8 million and peaks at 0.75 GB of memory rather than 1.4-1.7 GB. In tests across 240 sentences, Whisper transcription errors were nearly unchanged (4.69% versus 4.63%); UTMOS scores were also close, at 4.327 versus 4.383. The output is similar but not identical: some voiced sounds have a faint machine-like tone, and listening comparisons involved only the author. The model was trained and evaluated on Kokoro’s American and British English voices; other languages are untested, and Kokoro’s unchanged text side now accounts for most of the system’s parameters and CPU time.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.