---
id: "lib-2026-10-13-embedding-gemma-transcribe"
type: "library"
title: "Embedding Gemma Transcribe"
lang: "EN"
org: "haku.sh"
source_type: "Practitioner post"
date: "2026-10-13"
week: "2026-W42"
url: "https://nosov.ai/library/lib-2026-10-13-embedding-gemma-transcribe/"
original_url: "https://eg2.haku.sh/"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "A one-page product note from an independent builder with concrete speed and error figures from its own tests; a credible idea, but self-reported and without a published evaluation."
summary_basis: "full-text"
original_minutes: "2"
tags: ["speech recognition", "small models", "local inference"]
---

# Embedding Gemma Transcribe

Tier B (Useful, limited): A one-page product note from an independent builder with concrete speed and error figures from its own tests; a credible idea, but self-reported and without a published evaluation.

Summary based on the full text.

## In short

A developer at haku.sh turned Google's EmbeddingGemma 2, a multimodal embedding model never trained for speech-to-text, into an on-device transcriber for Macs. The audio encoder is left frozen. A small trained "reader" maps short slices of audio into the model's text space and picks the closest of about 229,000 English words, and a tiny extra model adds punctuation. It transcribes an hour of audio in 12.6 seconds on a base MacBook Air, at about 5% word error rate.

## Key numbers

- 1 hour in 12.6 s: transcription time for an hour of audio on a base M3 MacBook Air with 8 GB (288x real time)
- about 5%: word error rate, against 2-3% for leading speech models
- under two hours on 8×H100: training time for the reader, the only trained component
- ~229,000: English words the reader matches each audio slice against

## What it reports

- Speed: 288x real time on a base M3 MacBook Air (Neural Engine), against about 150x for IBM Granite Speech 5.0 TurboCTC on the same machine.
- Published figures for NVIDIA Parakeet variants on newer Macs range from 156x to 480x; the author calls this a rough guide, not a like-for-like ranking.
- Word error rate about 5%, against 2-3% for leading models; the author expects a better-trained reader to close the gap.
- Training: under two hours on 8×H100 on openly licensed English speech.
- Free macOS app, 715 MB of models on first launch; audio stays on the device.

## How they got it

The author's own speed tests on one laptop for the first two rows; other rows copied from FluidAudio's published benchmarks on different hardware and audio. No test set, error breakdown or code is given on the page.

## Limits

- No named test set for the word error rate, so the 5% figure cannot be checked.
- Speed comparisons mix the author's tests with third-party figures on different machines.
- English only; per-word matching against a fixed vocabulary will struggle with names and rare terms.

## Why it matters for companies

Private, local transcription of meetings and interviews is getting cheap enough to run on standard laptops, which removes one reason to send audio to a cloud service. The technique of reading another modality out of a shared embedding space with a small adapter may also cut the cost of building narrow tools. At 5% error, it suits drafts and search, not records that need to be exact.

Original: https://eg2.haku.sh/