Google launches Gemini 3.5 Transcribe, an accurate speech-to-text model with voice editing

Google announced Gemini 3.5 Transcribe, its most advanced speech-to-text model to date, and made it available to developers through the Gemini API in Google AI Studio and the Gemini enterprise agent platform. The model is offered in two separate tracks: real-time streaming via Live API with gemini-3.5-transcribe-live, which delivers sub-second latency for interactive voice applications, and batch processing of recorded audio via Interactions API with gemini-3.5-transcribe, which includes speaker attribution and word-level timestamps. Google emphasizes that the model is built to plug directly into developer workflows, whether they are building voice agents, real-time captioning tools, or analytics pipelines for recorded calls.

According to measurements by Artificial Analysis, the model achieves an average Word Error Rate (WER) of 4.0 % in streaming and 2.6 % in non-streaming use, with strong performance in noisy environments and in recognizing alphanumeric entities such as codes and order numbers. On the FLEURS benchmark covering leading languages and locales, the results are 5.50 % WER streaming and 5.04 % without streaming. Google cites a 70-percentage-point improvement in final transcription latency compared with its previous model, Chirp 3. The metrics come from a single source (Artificial Analysis) and have not been validated by multiple independent tests; support for 3+ speakers is still experimental.

The headline innovation is “smart transcription” rather than pure accuracy. The model automatically applies self-corrections (for example, “We’ll meet on Tuesday, no, Wednesday”), removes filler words such as “uh” and “um”, and formats the text automatically. An additional capability is Function Calling; the model can delegate complex tasks like image generation or file analysis to other Gemini models via function calls, a feature currently available only in the Gemini app for macOS. Other highlights include a domain-specific vocabulary with unique spelling handling, automatic detection of over 85 languages with regional accents and dialects, and speaker attribution for up to three participants in recorded audio.

Gemini 3.5 Transcribe is already integrated into Google consumer products: the Gemini app and Android via the Rambler feature, Gboard, Antigravity, and Chrome. In Gboard on Android, Rambler converts spoken thoughts into well-formatted text, filters filler words, enables voice-only editing, corrects spelling, changes writing style, and allows text modifications using voice alone. Google describes this as a step toward more natural interaction that understands context, intent, and inline edits.

On paper, Gemini 3.5 Transcribe represents a clear jump over Chirp 3, with new capabilities, lower WER, and significantly reduced latency. As with any infrastructure provider, the true test will be integration: how the model performs in production code with dirty audio, shifting acoustics, and tight latency constraints. Google’s decision not to disclose model architecture, training data, or licensing details beyond the API leaves developers with limited scope for independent evaluation before deployment.