The project page describes the adaptation, macOS app and author metrics without reproducible training artifacts.
Independent project adapts EmbeddingGemma 2 for speech transcription
2026-10-13Frontier watchIGNORE
An independent macOS project adapts EmbeddingGemma 2's audio encoder for speech transcription by training a small reader and punctuation model. The project reports roughly 5% word error rate and one hour of audio processed in about 13 seconds on a base MacBook Air. It distributes a free app, but no reproducible training code, test set or independent comparison is exposed.
What this means for you
The demo is an inventive reuse of a multimodal embedding model, but an opaque app and author-only metrics do not support engineering adoption. Record the community signal and wait for reproducible artifacts.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Platform | macOS 14+ on Apple Silicon |
|---|---|
| Download | free app; 715 MB models downloaded on first launch |
| Author Results | about 5% word error rate and one hour of audio in about 13 seconds on a base MacBook Air |
| Training | small reader trained for under two hours on 8 H100 GPUs; base audio encoder unchanged |
| Limits | no public training code, test set or independent evaluation |
| Time Basis | public Hacker News announcement at 2026-10-13T11:27:14Z; project page exposes no separate publication timestamp |
Technical impact 2
Engineering value 2
Adoption signal 1
Maturity 2
Verification cost 4
Risk 3