Voice typing is useful until it misunderstands your words.
It may include every “um.” It can miss names. It may also struggle when you correct yourself halfway through a sentence.
Google wants to fix that with Gemini 3.5 Transcribe.
The company announced the new speech-to-text model on August 26, 2026. It is designed to turn natural speech into clean and properly formatted text.
It cleans up your speech
Gemini 3.5 Transcribe does more than write down every word you say.
According to Google, it can remove filler words and format the final text automatically.
It can also understand corrections made while speaking.
Imagine saying:
“Let us meet on Tuesday. Sorry, I mean Wednesday.”
The model should understand that Wednesday is the correct day. It will not treat both dates as equally important.
This could make voice typing feel more natural. You can speak freely without carefully planning every sentence first.
It can learn difficult words
Transcription tools often struggle with names and technical terms.
Gemini 3.5 Transcribe allows users to provide custom vocabulary. This can include product names, medical terms, legal language or uncommon spellings.
That may be useful for Ghanaian users. Local names and places are regularly misheard by speech-to-text systems trained mainly on Western voices.
However, Google has not published detailed accuracy results for every Ghanaian accent. Important names and figures should still be checked manually.
It supports live and recorded audio
Google offers two versions of the model.
The live version transcribes speech while the person is still talking. Google says it can respond with less than one second of delay.
This could be useful for:
- Live captions
- Voice assistants
- Customer-service tools
- Interviews
- Online meetings
The recorded version processes saved audio. It can transcribe meetings, telephone calls, interviews and voice recordings.
It also provides word-level timestamps. This makes it easier to locate the exact moment a word was spoken.
Google says it can identify up to three speakers in a recording. Support for more speakers is still experimental.
More than 85 languages
According to Google, Gemini 3.5 Transcribe can automatically detect and transcribe over 85 languages.
It can also follow a speaker who changes languages during a conversation.
Support does not mean that every language and accent will perform equally well. Background noise, microphone quality and speaking speed can still affect the result.
Always review a transcript before publishing it or using it as an official record.
Where can you use it?
Developers can access Gemini 3.5 Transcribe through the Gemini API in Google AI Studio.
Some of its features are also appearing inside Google products.
On supported Android devices, Gboard’s Rambler feature can turn spoken thoughts into formatted writing. Users can correct spellings and change the writing style with voice instructions.
Google also says similar talk-to-type functionality is coming to Chrome. It will allow users to dictate replies, posts and prompts inside web text boxes.
The Gemini app on macOS can combine voice commands with information visible on the screen. Google says users can ask it to summarise files, rewrite text or activate other Gemini features.
These features may not reach every device or country at the same time.
How accurate is it?
Google cites results from Artificial Analysis showing an average word error rate of 4 percent for live transcription.
The recorded version achieved 2.6 percent in the cited testing.
A lower error rate means fewer transcription mistakes.
Google also says the new model can complete final transcriptions 70 percent faster than its previous Chirp 3 model.
These results are encouraging. However, real performance will depend on the language, accent, recording environment and device being used.
Delppy’s take
This could be very useful for creators.
It may reduce the time spent cleaning interviews, writing captions and converting voice notes into articles.
Students could use it to organise recorded lessons. Journalists could prepare interview drafts faster. Businesses could create meeting notes and call transcripts.
The real test for us will be how well it understands Ghanaian accents and names.
AI should handle the first draft. A human should still check the final transcript.
You can read Google’s official Gemini 3.5 Transcribe announcement for the complete technical details.
You can also explore our latest AI news and guides for more practical updates.
We go talk more chale.
