Written by David McMahon
Google has introduced Gemini 3.5 Transcribe, a speech-to-text model positioned less as a basic dictation engine and more as a layer for voice-driven workflows. Its August 26 announcement says the model turns raw audio into polished, formatted text while dealing with background noise, specialized vocabulary and spoken self-corrections. It is being offered in public preview through separate real-time and prerecorded-audio interfaces, a product choice that acknowledges the very different requirements of live conversation and after-the-fact analysis.
The split is important. The Live API exposes gemini-3.5-transcribe-live for continuous, bidirectional streaming with sub-second latency, aimed at interactive voice applications. The Interactions API exposes gemini-3.5-transcribe for recorded meetings, call logs and other files, with speaker attribution and word-level timestamps. A single model family can therefore serve captioning, voice agents and post-call analysis, but the data, latency and error-tolerance thresholds vary sharply among those uses.
Google’s product claims go beyond raw recognition. The model can smooth disfluencies, identify self-corrections, remove filler words and format output; it also supports custom vocabulary, automatic detection across more than 85 languages, and up to three-speaker attribution in prerecorded audio. Those features are valuable because many downstream AI tasks start with a transcript. If a transcript is wrong, an accurate-looking summary, action list or CRM update can still be wrong. Improving the source layer has compounding value across the rest of the workflow.
The same cleanup capability creates a governance issue. Verbatim transcription and intelligible transcription are not identical products. Removing an “um” is usually harmless, but resolving a self-correction, normalizing a technical term or polishing an ambiguous phrase can alter a record’s meaning. In legal, clinical, financial or investigative settings, users need to know whether they are reading the original utterance, a machine-cleaned version or a later summary. The reliable implementation pattern is to preserve the audio, retain timestamps, label transformations and give a reviewer a fast way to inspect the disputed segment.
Google cites Artificial Analysis measurements of 4.0% average word error rate for streaming and 2.6% for non-streaming use cases, as well as a 70% improvement in time to final transcription versus Chirp 3. These are useful benchmarks, but they should be understood as reported performance measurements, not a substitute for testing a deployment’s own accents, audio quality, jargon, speaker overlap and privacy constraints. A low aggregate error rate can still hide critical mistakes in names, numbers, medication terms, addresses or product codes.
The product also connects transcription to action. Google says function calling can let the model delegate tasks such as image generation and file analysis to other Gemini models, and notes integrations across the Gemini app, Android input, AI Studio and enterprise tooling. This is where the announcement becomes more significant than a model benchmark: speech is becoming an interface for structured work, not merely a format to be converted into text.
The adoption question is therefore not whether people will talk to software. They already do. It is whether organizations can define when automated cleanup is acceptable, when human confirmation is mandatory, and how an employee can correct the record without friction. Gemini 3.5 Transcribe gives developers more capable building blocks. The most durable implementations will be the ones that make accuracy, provenance and correction as visible as speed.
