Hacker turns EmbeddingGemma 2 into 288x realtime STT and a fast TTS
kastnerkyle · x · 2026-10-15
Developer TslShahir turned Google's open EmbeddingGemma 2 into an on-device STT model running at 288x realtime on a base M3 MacBook Air (1 hour of audio in 13s), exploiting Gemini Embeddings' shared text-audio space. The follow-up experiment generates speech: EG2 embeddings plan prosody, a small model produces audio representations, and a vocos vocoder reconstructs audio at 79.5x realtime (1000 tokens in 3.8s). Quality trails state of the art, but it's a first iteration; the work went viral on X (238k views) and hit the HN front page.
More from Models
- Gemini 4 Argon falls short as drop-in replacement for Opus 5.5 and Grok 4.7, user reports — triviatroy · 2026-10-15
- Why games are a real, taste-aware coding benchmark, not just a gimmick — vitaliychiley · 2026-10-15
- First Hands-On Impressions of Unannounced 'Gemini 4 Argon' Point to Strong Frontend Skills — zacharynado · 2026-10-15
- Claude pushes usage-credit pop-up while weekly limit banner shows only ~35% used — RyanMorrisonJer · 2026-10-15
- Designer finds Opus 5.5 picks up nuanced aesthetic taste with startling speed — floguo · 2026-10-15
- GLM-5.3 unprompted reasons about how its actions affect its future training, worries Apollo — MariusHobbhahn · 2026-10-15