Dev turns EmbeddingGemma 2 into an on-device STT model, transcribing 1 hour of audio in 13 seconds
solyarisoftware · x · 2026-10-13
Developer @TslShahir turned Google's open embedding model EmbeddingGemma 2 into a speech-to-text model, and it proved to be one of the fastest STT models you can run on device: on a base M3 MacBook Air it transcribes an hour of audio in about 13 seconds — 288× real time.
The idea traces back to Gemini Embedding 2 in March: such models place text, images, audio and video into one shared embedding space, so a spoken word lands next to the written word — meaning you could in theory read a transcript straight out of the embedding model. The author couldn't test it while it was closed, but tried once EmbeddingGemma 2 was released as an open model last week.
More from Models
- Dev says Codex is now 'nearly unusable': 4x slower than Opus 5.5 with worse results — TJLarkin23 · 2026-10-13
- oQe Quantization Update Cuts Code KL Divergence by Up to 18% Across 69 Models — huggingface · 2026-10-13
- Allen AI's TutorMoments Benchmark Finds LLM Tutors Over-Help Students — allen_ai · 2026-10-13
- DeepSeek's flash endpoint may be quietly testing a new model, users speculate v4.1-pro — teortaxesTex · 2026-10-13
- 3D game demos hint at AI creativity that transfers to CAD, docs and UI tasks — Angaisb_ · 2026-10-13
- subquadratic's New Model Beats Jev in 6 of 7 Production Use Cases Before Launch — ryanallred · 2026-10-13