Embedding Gemma 2-assisted speech generation reportedly reaches 79.5x real-time speed

A developer says their speech-generation experiment uses Embedding Gemma 2’s semantic word embeddings to plan prosody, a small model to generate an audio representation and a Vocos vocoder to reconstruct speech. They report producing 1,000 tokens in 3.8 seconds—79.5x real-time speed—on a base M3 Air, while noting the quality is not close to the best in the industry.