{
  "id": 14188291,
  "title": "An hour of audio transcribed in about 13s on MacBook Air",
  "url": "https://urgent.news/2026/10/13/an-hour-of-audio-transcribed-in-about-13s-on-macbook-air",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-13T11:27:14.000Z",
  "source": {
    "name": "Hacker News",
    "slug": "hacker-news",
    "url": "https://eg2.haku.sh/"
  },
  "original_language": "en",
  "account": "Google's EmbeddingGemma 2 model has been repurposed into a speech-to-text system, allowing an hour of audio to be transcribed in just 13 seconds on a standard MacBook Air, all performed locally on the device. This model is unique as it processes various media types—text, images, audio, and video—within a shared framework. By inputting spoken words, the model can generate corresponding text, demonstrating its ability to interpret audio inputs.\n\nThe model was trained using English speech data sourced from openly licensed materials, with the entire training process completed in under two hours on an 8×H100 GPU setup. Importantly, the base model did not originally possess speech-to-text capabilities. The transcription occurs in real-time, processing one second of audio per second of processing time. The findings presented include tests conducted on a MacBook Air equipped with an M3 chip, as well as comparisons with data from FluidAudio on more powerful, newer Mac models and varied audio inputs. These results are provided as a rough guide rather than a definitive ranking.\n\nCurrently, the model achieves a word error rate of approximately 5%, which is somewhat higher than the 2-3% error rate observed in leading speech recognition technologies. However, the creators anticipate that with further refinement, the model's performance could approach that of top-tier speech recognition systems. The transcription process involves simply dropping an audio file onto the device or, alternatively, using a dedicated key combination to activate voice dictation within any application. Crucially, all processing happens on the Mac's Neural Engine, ensuring that the audio data remains stored locally on the device.",
  "summary": null,
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}