Gemini 3.5 Transcribe cleans up speech as it transcribes

Google’s new model transcribes live or recorded audio, corrects disfluencies, identifies speakers, and supports more than 85 languages.

Transcribing a conversation without preserving every “um,” repetition, or abandoned sentence: Google is introducing Gemini 3.5 Transcribe, a speech recognition model designed to produce corrected, punctuated, and structured text directly from raw audio.

The approach goes beyond word-for-word transcription. In Smart mode, the model recognizes corrections made while speaking. Someone saying, “Let’s meet Tuesday—no, Wednesday” receives a sentence that only retains Wednesday. Hesitations, filler words, and repetitions can be removed, while lists, numbers, currencies, and punctuation are automatically formatted.

A Verbatim mode remains available when fidelity to the spoken words matters most, such as for interviews, linguistic research, or legal documents. The distinction is important: Smart mode produces more readable text, but it also modifies the speaker’s original phrasing.

Gemini 3.5 Transcribe is available in two versions. `gemini-3.5-transcribe-live` processes a continuous audio stream through the Live API and returns text with subsecond latency. It targets voice assistants, live captioning, and voice-controlled interfaces. Streaming sessions are currently limited to ten minutes.

The `gemini-3.5-transcribe` model processes recorded files through the Interactions API. It is intended for meetings, interviews, calls, and audio archives, and can provide timestamps for every word. Google warns that enabling this option may slightly reduce overall transcription accuracy.

Recorded audio processing can also distinguish between speakers. Up to eight voices can be identified and assigned separate labels, although Google still considers conversations involving more than three speakers experimental. This capability is unavailable in the live version, which also does not provide word-level timestamps.

The model automatically detects more than 85 languages and regional variants. It can follow language switches within a sentence or conversation without prior configuration. Developers can still specify the expected language to improve recognition.

Custom vocabulary settings allow users to add up to 1,000 terms, although Google recommends limiting the list to around 100 for the strongest results. The feature is intended to improve the transcription of proper names, brands, acronyms, technical references, and industry-specific terminology.

Google says the model performs better in noisy environments and recognizes alphanumeric sequences such as order numbers and postal codes more accurately. According to Artificial Analysis measurements cited by the company, its average Word Error Rate reaches 4% for live transcription and 2.6% for recorded audio. The time required to obtain a final transcript is reportedly 70% shorter than with Chirp 3, Google’s previous speech recognition model.

These results still depend on the language, accent, recording quality, and surrounding context. On the multilingual FLEURS benchmark, Google reports error rates of 5.50% in streaming mode and 5.04% for recorded audio. A general average therefore does not guarantee the same accuracy across every language or situation.

The model can also route certain requests to other Gemini systems. A voice command could trigger a search, file analysis, or another action inside a compatible application. This capability is used in the Gemini app for macOS, where transcription can incorporate on-screen context with the user’s permission.

Gemini 3.5 Transcribe is available in public preview through the Gemini API and Google AI Studio. It is also integrated into Rambler in Gboard on selected Android devices, the English-language Gemini app for macOS, and Google Antigravity. A dictation feature for text fields in Chrome is expected later.

On the API’s paid tier, Google estimates the combined cost of transcribing a recorded file at approximately $0.005 per minute. Live transcription costs around $0.009 per minute when audio input and text output are combined. A free tier is available, but submitted data may be used to improve Google’s products. The company says this does not apply to paid API requests.

Google is therefore moving beyond simply converting speech into text. Gemini 3.5 Transcribe attempts to understand what the speaker intended to say and deliver a version that is already usable. That could save time for dictation and voice interfaces, provided users choose carefully between a faithful transcript and text edited by the model.