Google Releases Gemini 3.5 Transcribe Models

Google describes the new transcription behavior as more than verbatim speech recognition. Gemini 3.5 Transcribe can remove filler words such as "ums" and "ahs," process self-corrections such as a speaker changing a meeting day, and automatically format the resulting text. Google also states that the models are designed to work in noisy environments and with complex vocabulary.

That behavior creates a material distinction between a raw transcript and an edited transcript. Teams using speech data for compliance review, legal records, clinical documentation, or conversational analytics commonly need to retain the original audio and establish clear review policies, because removal of disfluencies and resolution of self-corrections can change the textual representation of what was said.

Google states that gemini-3.5-transcribe-live is available through its Live API, while gemini-3.5-transcribe is available through the Gemini API. The company also describes function calling in the Live API, allowing the model to delegate tasks such as image generation and file analysis to other Gemini models.

Google cites Artificial Analysis measurements of average word error rates of 4.0% for streaming use cases and 2.6% for non-streaming use cases. Those figures are useful reference points, but practitioners evaluating production speech systems typically need task-specific testing across accents, acoustic conditions, domain terminology, speaker overlap, and punctuation requirements. Aggregate word error rate alone does not measure whether edited output preserves a workflow's required audit trail.

The release extends Gemini Audio from consumer voice features into developer workflows where latency, diarization, vocabulary handling, and transcript normalization can be evaluated as separate system requirements.