Google 发布 Gemini 3.5 Transcribe，实现 AI 驱动的语音转文字

在等待 Gemini 3.5 Pro 发布的同时（或许只是徒劳等待），Google 正在推出 3.5 系列中的另一款模型。该公司宣布推出 Gemini 3.5 Transcribe，这是一款旨在简化语音输入的 AI 模型：它会删除“嗯”以及自我修正等内容，并输出经过润色的 AI 文本。这一模型已经为 Pixel 11 上 Gboard 的“Rambler”功能提供支持，但很快将在整个 Google 生态系统中出现。

据 Google 介绍，Gemini 3.5 Transcribe 的速度和准确性都显著优于此前名为 Chirp 3 的语音转文字引擎。从语音到最终转录文本，新 AI 模型的速度预计将提升约 70%，实时语音错误率则降至 5.5%。这相比 Chirp 3 的表现仅略有提升，Google 对 Chirp 3 的测量结果为 7.32%。不过，使用语音输入时修正错别字相当麻烦，因此这里的任何改进都有益处。

新模型不仅更擅长听清你说出的词，还据称更能抓住你真正想表达的核心意思。在你说话时，Gemini 3.5 Transcribe 可以删除那些会让意识流表达变得杂乱的“嗯”和“呃”。如果你需要纠正自己，它也能实时编辑文本；同时还可参考你提供的自定义词汇表，以处理“专业术语”。这些功能支持 85 种语言，并可处理最多三位说话人的预录音频。

当然，其缺点在于，你需要依赖 AI 准确理解你讲话的大意。对于较短的文本块，该模型似乎很擅长清理前后不一致之处和口头表达中的磕绊（根据我对 Rambler 的测试），但 AI 在技术上确实会改变你说过的话的措辞，这可能并不适合所有场景。

While we wait (possibly in vain) for Gemini 3.5 Pro to launch, Google is releasing a different model in the 3.5 branch. The company has announced Gemini 3.5 Transcribe, an AI model designed to streamline voice input by editing out “ums” and corrections, outputting polished AI text. This model already powers the Gboard “Rambler” feature on the Pixel 11, but it’s about to appear throughout the Google ecosystem.

According to Google, Gemini 3.5 Transcribe is much faster and more accurate than its previous voice-to-text engine, known as Chirp 3. The new AI model should be about 70 percent faster from voice to final transcribed text, and the live-speech error rate has dropped to 5.5 percent. That’s only a little better than Chirp 3, which Google measures at 7.32 percent. Still, it’s a pain to fix typos when you’re using voice input, so any improvement here is beneficial.

The new model isn’t just better at hearing the words you say; it’s also supposedly better at getting to the heart of what you meant to say. As you speak, Gemini 3.5 Transcribe can remove the awkward “ums” and “uhs” that clutter your stream of consciousness. It can also edit text on the fly if you need to correct yourself and refer to your provided custom vocabulary to handle “specialized jargon.” All this works in 85 languages and with up to three speakers in pre-recorded audio.

The drawback, of course, is that you’re relying on the AI to accurately get the gist of your speech. For short blocks of text, the model seems good at cleaning up inconsistencies and verbal stumbles (based on my testing with Rambler), but the AI does technically change the wording of what you said, and that may not be appropriate for all situations.