Google has unveiled its most ambitious offering yet in the field of speech recognition. Called Gemini 3.5 Transcribe , it's not your typical speech-to-text converter: it promises to understand what we say, clean it up, and return polished text, free of filler words and with correct punctuation. The company presents it as its most accurate model to date, and initial data suggests a significant improvement, especially in noisy environments or with complex accents.
The big news is that the system doesn't just copy what it hears. If you say "let's meet on Tuesday... no, on Wednesday," the final result will be "let's meet on Wednesday." It also removes "uh" and "em," formats numbers and postal codes, and even recognizes specialized vocabulary if provided with a custom list. All of this is done with a speed that Google claims is 70% faster than its previous engine, Chirp 3 , and with support for more than 85 languages.
A transcript that understands what you mean
The key to Gemini 3.5 Transcribe lies in its intelligent transcription. The model is trained to grasp the speaker's intent, not just individual words. This is evident in details like autocorrections: if you make a mistake mid-sentence and correct yourself, the system retains the final version. It can also distinguish up to three speakers in a pre-recorded audio file and assign timestamps to each speaker, which is very useful for meetings, interviews, or call analysis.
In terms of accuracy, Google's data shows a word error rate (WER) of 4% for streaming and 2,6% for pre-recorded audio, according to measurements from Artificial Analysis. This means that, in practice, errors are reduced to fewer than three words per hundred under normal conditions. The company also highlights its performance in noisy environments, where other models often struggle.
Another advantage of the model is its custom vocabulary. You can upload a list of proper nouns, acronyms, or technical terms, and the system will recognize them without you having to spell them out. This is especially useful for doctors, lawyers, or journalists who work with specialized jargon. Furthermore, the model automatically detects the language and can switch between them mid-conversation, something that was previously a major headache.
Where it is available and what products use it
Gemini 3.5 Transcribe isn't just a lab experiment. Google has already integrated it into several of its products. The most visible example is Rambler, Gboard's dictation feature for Android , which now uses this model to offer cleaner transcriptions on Android keyboards . It's also present in the Gemini app for macOS, where it allows users to dictate voice commands that are combined with the screen context to summarize files, generate images, or reuse text between applications.
For developers, the model is available in public preview through the Gemini API, both in Google AI Studio and Antigravity. It can also be tested on the Gemini Enterprise Agent Platform, the company's enterprise platform. With these tools, technical teams can create voice agents, real-time captioning systems, or call analytics tools without relying on external services.
The next stop will be Chrome. Google has confirmed it's working on integrating the model into the browser, so users can dictate text in any web field without switching applications. This would open the door to composing emails, filling out forms, or interacting with Gemini directly using voice, although there's still no specific release date.
Preview version pricing and limitations
The developer version has its own rules. The Live API, designed for real-time streaming, allows sessions of up to ten minutes and does not include speaker identification. The Interactions API, for pre-recorded audio, supports files up to one hour long and does offer diarization with up to three voices, although attribution to more than three speakers is considered experimental. Regarding pricing, Google estimates a combined cost of $0,009 per minute for streaming and $0,005 per minute for files , although these are indicative calculations based on token consumption.
It's important to note that availability varies by country and language. Rambler, for example, is only available in certain markets, and the Gemini app for macOS currently only works in English. Google hasn't provided details on when support will expand to more regions, but it seems likely to be within a few months.
On the competitive front, Gemini 3.5 Transcribe comes with a strong argument: integrated automatic correction. Services like OpenAI's Whisper or AssemblyAI offer good accuracy, but they lack the editing layer that removes filler words and formats the text. However, experts point out that human oversight remains necessary in contexts where literal fidelity is critical, such as legal statements or medical reports.
What does this mean for the average user?
For those who dictate messages, notes, or emails from their mobile phone, the difference is noticeable from the first use. Transcriptions arrive cleaner, without those "uh"s and "um"s that then have to be manually deleted, and with the punctuation correctly placed. It's also appreciated that the system understands corrections on the fly, which saves time and prevents frustration.
In the professional sphere, this tool can be a game-changer for journalists, researchers, or assistants who work with recordings. The ability to identify who speaks and when in a meeting drastically reduces editing time. And for customer service teams, post-call analysis becomes much simpler with accurate and labeled transcripts.
However, it's worth remembering that the model is in preview and that some features, such as diary entries with more than three voices, are still experimental. Google has promised improvements in the coming months, but for now, the recommendation is to always review the transcripts before using them officially.
With the arrival of Gemini 3.5 Transcribe, Google demonstrates that voice can be as powerful an interface as a keyboard, provided the technology is up to par. The combination of speed, accuracy, and automatic editing makes it a very attractive tool for both individual users and businesses. It remains to be seen how it will perform in the real world, but initial impressions are promising.
