EmbeddingGemma 2 Converted into Speech‑to‑Text Model
embeddings gemma google multimodal speech
| Source: HN | Original article
Developers repurposed Google’s multimodal EmbeddingGemma 2, originally built for cross‑modal embeddings, into a functional speech‑to‑text model.
A developer has repurposed Google’s open‑source EmbeddingGemma 2 into a full‑speed speech‑to‑text engine, releasing the tool as “Embedding Gemma Transcribe,” a free macOS app. The original model, a sub‑1 billion‑parameter multimodal encoder built on the Gemma 4 architecture, maps text, images, audio and video into a shared 768‑dimensional embedding space. By leveraging its native audio embeddings, the creator re‑engineered the model to output transcriptions instead of vectors, achieving an hour‑of‑audio conversion in roughly 13 seconds on a base‑spec M3 MacBook Air.
The hack matters because it showcases how a compact, on‑device embedding model can be transformed into a high‑throughput, privacy‑preserving speech recogniser without relying on cloud services. EmbeddingGemma 2’s Apache 2.0 licence and modest hardware requirements make it attractive for developers seeking fast, local transcription for personal or enterprise workflows. The speed demonstrated—over 270× real‑time—suggests that similar repurposing could bring competitive on‑device speech capabilities to a broader range of consumer hardware, challenging larger, server‑based solutions.
Observers should watch whether Google or the broader open‑source community formalises this approach, perhaps by publishing a dedicated speech‑to‑text variant or integrating the technique into its upcoming multimodal tools. Further benchmarks on other Apple silicon chips, Windows laptops or edge devices will indicate how widely the method scales. Finally, the move may spur additional experiments that turn other multimodal embeddings—such as those used for cross‑modal retrieval—into specialised on‑device AI services, accelerating the shift toward locally run, low‑latency intelligence.
Sources
Back to AIPULSEN