We turned EmbeddingGemma 2 into a speech-to-text model
- ID
- 34918
- Status
- summarized
- Published
- 13 Oct 2026, 7:27 PM
- Fetched
- 14 Oct 2026, 12:55 AM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://eg2.haku.sh/
- Source URL
- https://hnrss.org/best
Summary
- Score
- 7.5
- Created
- 14 Oct 2026, 12:55 AM
- Tags
- Audience
- developersai_ml_learnersvibe_coderssaas_founders
What happened
A developer built 'Embedding Gemma Transcribe', which repurposes Google's EmbeddingGemma 2 multimodal embedding model into a speech-to-text tool by training only a small 'reader' that maps audio slices into the model's existing text space and picks the nearest of ~229,000 English words. It transcribes an hour of audio in about 12.6 seconds on a base M3 MacBook Air with 8 GB RAM (roughly 288x real time, running on the Neural Engine, fully on-device), versus ~150x for IBM Granite Speech 5.0 TurboCTC via MLX on the same machine. The tradeoff: word error rate is about 5% versus 2–3% for leading speech models, and the reader was trained in under two hours on 8xH100 using openly licensed English speech; the free Mac app (v1.6, macOS 14+, 715 MB model download) supports file input and hold-Right-Option dictation into any app.
Why it matters
If you handle sensitive audio — client calls, internal meetings, customer interviews — this gives you a free, offline, on-device option where audio never leaves the Mac, at 288x real time, but you accept roughly double the error rate of leading cloud speech models (5% vs 2–3%). It is English-only (229,000-word vocabulary), so Malay, Mandarin, or Tamil transcription is not covered by this build — do not assume it works for local-language content. The published comparison rows against Granite, Parakeet, and Phonon-2 use larger Macs and different audio, so treat them as a rough guide, not a like-for-like benchmark.
Discussion angle
Is retraining a tiny 'reader' on top of a frozen embedding space a real alternative to dedicated ASR models, or just a clever demo — and at 5% WER, what workloads would you actually trust it with?