@aigclink: Someone turned Google's EmbeddingGemma 2 into a speech recognition model, achieving an astonishingly fast on-device inf…

X AI KOLs Timeline News
speech-recognition embeddings on-device multimodal retrieval gemma

Summary

有人将谷歌的 EmbeddingGemma 2 改造为语音识别模型，通过在冻结参数的嵌入模型上训练轻量级 Reader，把语音转写从生成问题变成检索问题，在 M3 MacBook Air 上实现 288× 实时速度（1 小时音频约 13 秒转写），WER 约 5%，可能重塑端侧 STT 市场。

Someone turned Google's EmbeddingGemma 2 into a speech recognition model, achieving an astonishingly fast on-device inference speed—insanely quick M3 MacBook Air, 1 hour of audio transcribed in about 13 seconds, processing speed 288× WER 5%, leading STT at 2%～3%. If subsequent reader iterations can push WER below 3%, the dedicated on-device STT market is probably going to get reshuffled Traditional STT requires training a dedicated voice→text generation model, but today, EmbeddingGemma 2 can already map text, images, audio, and video into the same semantic space, bringing the audio "hello" and the text "hello" closer together in vector space So netizens, without modifying EmbeddingGemma 2's parameters, just trained a lightweight Reader on top of it: after encoding audio slices into vectors, they directly retrieve the nearest words in the model's text space, turning transcription from a generation problem into a retrieval problem, essentially "hacking" it into a speech-to-text model The current 5% WER makes it practically viable for scenarios with relatively high error tolerance, like organizing meeting recordings, interview transcription, voice notes, video subtitle drafts, etc. From @TslShahir #EmbeddingGemma2 #STT #语音转文字
Original Article
View Cached Full Text

Cached at: 10/13/26, 03:47 PM

Someone turned Google’s EmbeddingGemma 2 into a speech recognition model, achieving an astonishingly fast on-device inference speed—insanely quick

M3 MacBook Air, 1 hour of audio transcribed in about 13 seconds, processing speed 288×

WER 5%, leading STT at 2%～3%. If subsequent reader iterations can push WER below 3%, the dedicated on-device STT market is probably going to get reshuffled

Traditional STT requires training a dedicated voice→text generation model, but today, EmbeddingGemma 2 can already map text, images, audio, and video into the same semantic space, bringing the audio “hello” and the text “hello” closer together in vector space

So netizens, without modifying EmbeddingGemma 2’s parameters, just trained a lightweight Reader on top of it: after encoding audio slices into vectors, they directly retrieve the nearest words in the model’s text space, turning transcription from a generation problem into a retrieval problem, essentially “hacking” it into a speech-to-text model

The current 5% WER makes it practically viable for scenarios with relatively high error tolerance, like organizing meeting recordings, interview transcription, voice notes, video subtitle drafts, etc.

From @TslShahir

#EmbeddingGemma2 #STT #语音转文字

AIGCLINK (@aigclink): 刚刚谷歌开源了其最新Embedding模型：EmbeddingGemma 2，原生多模态，手机、端侧能跑

去年的EmbeddingGemma是纯文本嵌入，这代把代码、图像、视频、音频全部打通，都映射进同一个向量空间

上下文8K，上一代的4倍，支持100+语言，跟Gemma 4搭配可跑端侧多模态RAG

Similar Articles

An hour of audio transcribed in about 13s on MacBook Air

Hacker News Top

Google's EmbeddingGemma 2 embedding model has been repurposed into a speech-to-text tool that transcribes an hour of audio in about 13 seconds entirely on-device on a base MacBook Air, using a small trained reader on top of the frozen audio encoder.

Google EmbeddingGemma 2

Product Hunt

Google DeepMind发布了EmbeddingGemma 2，这是一个开源的轻量级多模态嵌入模型，可将文本、代码、图像、视频和音频映射到同一共享空间，740M参数可在设备端运行，并采用Apache 2.0许可。