We turned EmbeddingGemma 2 into a speech-to-text model

Google’s EmbeddingGemma 2 is an embedding model. We turned it into a speech-to-text model: an hour of audio in about 13 seconds on a base MacBook Air, entirely on device.

Google 的 EmbeddingGemma 2 是一个嵌入模型。我们将其转变为一个语音转文字模型：在基础款 MacBook Air 上，处理一小时音频仅需约 13 秒，且完全在设备端运行。

EmbeddingGemma 2 is a multimodal embedding model: text, images, audio and video in one shared space. Say a word out loud and its audio lands close to the written word.

EmbeddingGemma 2 是一个多模态嵌入模型：它将文本、图像、音频和视频置于同一个共享空间中。当你大声说出一个词时，其音频在空间中的位置会非常接近该词的书面形式。

So we tried reading words straight out of that space.

因此，我们尝试直接从该空间中读取词汇。

Only the reader is trained: under two hours on 8×H100, on openly licensed English speech. The base model was never trained for speech-to-text.

只有读取器经过了训练：在 8×H100 上使用开源许可的英语语音数据，训练时间不到两小时。基础模型从未针对语音转文字进行过训练。

× real time: seconds of audio per second of processing. Rows 1–2: our tests on the same base M3 MacBook Air. Rows 3–5: FluidAudio’s published benchmarks on larger, newer Macs and different audio. A rough guide, not a like-for-like ranking.

× 实时性：指每秒处理的音频秒数。第 1-2 行：我们在同一台基础款 M3 MacBook Air 上的测试结果。第 3-5 行：FluidAudio 在更大、更新的 Mac 机型及不同音频上发布的基准测试。这仅作为粗略参考，而非同类项的直接排名。

Its word error rate is about 5% right now, against 2–3% for leading speech models. We believe a better-trained reader can close that gap; that’s the next iteration.

其目前的词错误率约为 5%，而领先的语音模型为 2–3%。我们相信，经过更好训练的读取器可以缩小这一差距；这就是下一次迭代的目标。

Drop in an audio file, or hold Right Option, speak and release to dictate into any app. It runs on the Mac’s Neural Engine; your audio never leaves the machine.

拖入音频文件，或者按住 Right Option 键说话，松开后即可在任何应用程序中进行听写。它运行在 Mac 的神经网络引擎上；你的音频永远不会离开这台设备。