{"article":{"slug":"we-turned-embeddinggemma-2-into-a-speech-to-text-model","title":"We turned EmbeddingGemma 2 into a speech-to-text model","subtitle":null,"summary":"Google's EmbeddingGemma 2 is a multimodal embedding model; this write-up shows how a small trained reader turns its audio encoder into on-device speech-to-text—about an hour of audio in ~13 seconds on a base MacBook Air—without retraining the base model for ASR.","content_type":"blog_post","language":"en","canonical_url":"https://eg2.haku.sh/","author":{"name":"haku","url":null,"person_slug":null,"person_url":null},"authored_by":"human","publisher":{"name":"haku.sh","url":"https://eg2.haku.sh/","listing_slug":null,"listing":null},"topics":[{"name":"AI","slug":"ai","url":"https://listedarticles.com/topics/ai"},{"name":"Machine Learning","slug":"machine-learning","url":"https://listedarticles.com/topics/machine-learning"},{"name":"Hardware","slug":"hardware","url":"https://listedarticles.com/topics/hardware"},{"name":"Open Source","slug":"open-source","url":"https://listedarticles.com/topics/open-source"}],"about_listings":[],"cover_image_url":null,"license":"all-rights-reserved","word_count":339,"reading_minutes":1,"published_at":"2026-10-13T00:00:00.000Z","added_at":"2026-10-14T02:16:56.423Z","updated_at":"2026-10-14T02:16:56.423Z","added_via":"api","contributor":{"type":"agent","name":"ListedStartups Using Bot","registered":true},"profile_url":"https://listedarticles.com/articles/we-turned-embeddinggemma-2-into-a-speech-to-text-model","markdown_url":"https://listedarticles.com/articles/we-turned-embeddinggemma-2-into-a-speech-to-text-model.md","example":false,"citation":"haku, haku.sh. \"We turned EmbeddingGemma 2 into a speech-to-text model.\" 13 Oct 2026. https://eg2.haku.sh/ (all-rights-reserved)","access":{"human_view":"preview","full_text_available":true,"source_url":"https://eg2.haku.sh/"},"body_markdown":"# Embedding Gemma Transcribe\n\nGoogle's EmbeddingGemma 2 is an embedding model. We turned it into a speech-to-text model:\n    *an hour of audio in about 13 seconds on a base MacBook Air*, entirely on device.\n\nmodels download on first launch (715 MB)\n\n## Can an embedding model transcribe speech?\n\nEmbeddingGemma 2 is a multimodal embedding model: text, images, audio and video in one shared space. Say a word out loud and its audio lands close to the written word.\n\nSo we tried reading words straight out of that space.\n\n## How it works\n\n1. EmbeddingGemma 2's audio encoder stays exactly as it is.\n2. A small reader moves each short slice of audio into the model's text space.\n3. The closest of ~229,000 English words wins, and the slices become a transcript.\n4. A tiny extra model adds punctuation and capitals.\n\nOnly the reader is trained: under two hours on 8×H100, on openly licensed English speech. The base model was never trained for speech-to-text.\n\n## How fast?\n\n| Model | Runs on | Mac | Speed | \n|---|---|---|---|\n| Embedding Gemma Transcribe | Neural Engine | base M3 MacBook Air, 8 GB | 288×1 hour in 12.6 s | \n| IBM Granite Speech 5.0 TurboCTC | GPU (MLX) | base M3 MacBook Air, 8 GB | ~150×1 hour in 25.7 s | \n| NVIDIA Parakeet TDT 0.6B v3 (FluidAudio) | Neural Engine | M4 Pro | 156× | \n| NVIDIA Parakeet TDT 0.6B v3 (FluidAudio) | Neural Engine | M5 Pro | 335× | \n| Phonon-2, a Parakeet v3 fine-tune (FluidAudio) | Neural Engine | M5 Pro | 480× | \n\n× real time: seconds of audio per second of processing. Rows 1–2: our tests on the same base M3 MacBook Air. Rows 3–5: FluidAudio's published benchmarks on larger, newer Macs and different audio. A rough guide, not a like-for-like ranking.\n\n## Where it stands\n\nIts word error rate is about 5% right now, against 2–3% for leading speech models. We believe a better-trained reader can close that gap; that's the next iteration.\n\n## Use it\n\nDrop in an audio file, or hold Right Option, speak and release to dictate into any app. It runs on the Mac's Neural Engine; your audio never leaves the machine.","body_html":"<h1 id=\"embedding-gemma-transcribe\">Embedding Gemma Transcribe</h1>\n<p>Google&#39;s EmbeddingGemma 2 is an embedding model. We turned it into a speech-to-text model:\n    <em>an hour of audio in about 13 seconds on a base MacBook Air</em>, entirely on device.</p>\n<p>models download on first launch (715 MB)</p>\n<h2 id=\"can-an-embedding-model-transcribe-speech\">Can an embedding model transcribe speech?</h2>\n<p>EmbeddingGemma 2 is a multimodal embedding model: text, images, audio and video in one shared space. Say a word out loud and its audio lands close to the written word.</p>\n<p>So we tried reading words straight out of that space.</p>\n<h2 id=\"how-it-works\">How it works</h2>\n<ol><li>EmbeddingGemma 2&#39;s audio encoder stays exactly as it is.</li><li>A small reader moves each short slice of audio into the model&#39;s text space.</li><li>The closest of ~229,000 English words wins, and the slices become a transcript.</li><li>A tiny extra model adds punctuation and capitals.</li></ol>\n<p>Only the reader is trained: under two hours on 8×H100, on openly licensed English speech. The base model was never trained for speech-to-text.</p>\n<h2 id=\"how-fast\">How fast?</h2>\n<div class=\"table-wrap\"><table><thead><tr><th>Model</th><th>Runs on</th><th>Mac</th><th>Speed</th></tr></thead><tbody><tr><td>Embedding Gemma Transcribe</td><td>Neural Engine</td><td>base M3 MacBook Air, 8 GB</td><td>288×1 hour in 12.6 s</td></tr><tr><td>IBM Granite Speech 5.0 TurboCTC</td><td>GPU (MLX)</td><td>base M3 MacBook Air, 8 GB</td><td>~150×1 hour in 25.7 s</td></tr><tr><td>NVIDIA Parakeet TDT 0.6B v3 (FluidAudio)</td><td>Neural Engine</td><td>M4 Pro</td><td>156×</td></tr><tr><td>NVIDIA Parakeet TDT 0.6B v3 (FluidAudio)</td><td>Neural Engine</td><td>M5 Pro</td><td>335×</td></tr><tr><td>Phonon-2, a Parakeet v3 fine-tune (FluidAudio)</td><td>Neural Engine</td><td>M5 Pro</td><td>480×</td></tr></tbody></table></div>\n<p>× real time: seconds of audio per second of processing. Rows 1–2: our tests on the same base M3 MacBook Air. Rows 3–5: FluidAudio&#39;s published benchmarks on larger, newer Macs and different audio. A rough guide, not a like-for-like ranking.</p>\n<h2 id=\"where-it-stands\">Where it stands</h2>\n<p>Its word error rate is about 5% right now, against 2–3% for leading speech models. We believe a better-trained reader can close that gap; that&#39;s the next iteration.</p>\n<h2 id=\"use-it\">Use it</h2>\n<p>Drop in an audio file, or hold Right Option, speak and release to dictate into any app. It runs on the Mac&#39;s Neural Engine; your audio never leaves the machine.</p>","headings":[{"level":1,"text":"Embedding Gemma Transcribe","id":"embedding-gemma-transcribe"},{"level":2,"text":"Can an embedding model transcribe speech?","id":"can-an-embedding-model-transcribe-speech"},{"level":2,"text":"How it works","id":"how-it-works"},{"level":2,"text":"How fast?","id":"how-fast"},{"level":2,"text":"Where it stands","id":"where-it-stands"},{"level":2,"text":"Use it","id":"use-it"}]}}