Google’s new AI transcription edits out your ‘ums’ and ‘ahs’
Google has rolled out updated Gemini Audio capabilities built on new Gemini 3.5 models, adding a dedicated transcription tool alongside upgraded live voice features. The launch matters because it targets long-standing weaknesses in AI dictation and voice assistants, such as coping with background noise, interrupted speech and messy filler words, while also expanding language support to over 85 tongues and improving handling of specialist terminology.
The centrepiece is Gemini 3.5 Transcribe, a new model Google says markedly improves on its previous Chirp 3 system, particularly for multilingual accuracy and error rates. It can automatically strip out filler words like "um" and "uh", format text, accept custom vocabulary for jargon or unusual spellings, and identify up to three speakers in pre-recorded audio with word-level timestamps. It launches alongside Gemini 3.5 Live and Live Experimental, which improve real-time interruption handling and visual processing, with the latter also narrating its reasoning step by step; the updates begin rolling out today for macOS Gemini app users and Android's Rambler dictation feature, plus developer preview access via the Gemini API, with Chrome support promised soon.
- Google launches Gemini 3.5 Transcribe, a new AI transcription model
- It removes filler words like "um" and automatically formats text
- Rolling out now on macOS and Android, Chrome support coming soon
New here? Start with this
Voice transcription turns spoken audio into written text. Speech often includes pauses, repeated words and fillers such as “um” and “uh”, which can appear in a transcript unless a system is designed to remove or format them.
Google’s Gemini is a family of AI models, and Chirp 3 was its earlier speech recognition system. The new audio tools are intended for uses ranging from dictation to live voice conversations, where background noise, interruptions, different languages and specialist terms can affect how well software understands speech.
Transcription can save time for people creating notes, captions or written records, but changes such as removing fillers can affect how closely the text reflects the original speech. Speaker labels and timestamps can help readers connect parts of a transcript to the audio.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Automatic removal of filler words represents a practical enhancement that serves genuine user needs. Most people want clean, readable transcripts for professional documents, meeting notes and accessibility purposes; filler words add no semantic content and clutter text. This feature delivers what users actually want from transcription services, and like other formatting improvements, it represents genuine progress in making AI voice tools more useful and fit for purpose.
The case against
Transcription systems should provide faithful records of what was actually spoken, not edited versions selected by machines. When software automatically removes content without explicit user control, it alters the source material and raises questions about authenticity and speaker agency, particularly in professional, legal or research contexts where precision matters. Filler words remain part of actual speech, and their removal without consent means users cannot later analyse or recover what was genuinely said.