Google has launched Gemini 3.5 Transcribe, a dedicated speech-to-text model designed for real-time streaming, automated disfluency cleanup, and agentic task delegation. The release introduces two API interfaces alongside integration across Google developer tooling and consumer operating system surfaces.
Dual API Architecture for Live and Batch Audio
Gemini 3.5 Transcribe is split into two operational endpoints tailored for distinct latency profiles:
- Real-time streaming (
gemini-3.5-transcribe-live): Accessible via the Gemini Live API, this variant provides continuous bidirectional streaming with sub-second latency for interactive voice applications and live captioning. - Pre-recorded audio processing (
gemini-3.5-transcribe): Operating through the Interactions API, the batch variant transcribes stored audio files, meetings, and call logs with word-level timestamps and automated speaker diarization for up to three speakers.
Integrated Disfluency Filtering and Custom Dictionaries
Standard speech recognition stacks often require a secondary large language model pass to remove filler words, format punctuation, and resolve speech restarts. Gemini 3.5 Transcribe handles these transformations natively during acoustic decoding.
The model automatically detects and strips conversational disfluencies ("ums" and "ahs") while resolving spoken mid-sentence corrections (such as shifting scheduled dates or addresses) into clean final text. In addition, developers can inject custom vocabulary sets at inference time to preserve proprietary entity names, technical jargon, and alphanumeric identifiers such as postal codes or tracking numbers.
Multilingual coverage spans more than 85 languages with automated language detection, accent normalization, and code-switching support.
Accuracy and Latency Benchmarks
According to benchmark measurements reported by Artificial Analysis, Gemini 3.5 Transcribe achieves:
- Word Error Rate (WER): 4.0% in streaming mode and 2.6% in non-streaming batch mode.
- FLEURS Multilingual Benchmark: 5.50% streaming WER and 5.04% batch WER across primary language evaluation suites.
- Latency Reduction: A 70% decrease in time-to-final-transcription compared to Google's previous Chirp 3 speech architecture.
Tool Calling and Agentic Voice Workflows
Beyond passive speech capture, Gemini 3.5 Transcribe is wired into Google's function-calling ecosystem. In developer environments like Google Antigravity and the Gemini macOS desktop client, the transcription layer operates alongside active screen context and file trees.
Spoken prompts can trigger background function execution across other Gemini models, routing complex requests such as image generation, local document summarization, or code editing directly from raw audio streams without separate orchestration middleware.
Availability and Deployment
Gemini 3.5 Transcribe is available in public preview for developers through Google AI Studio, Google Antigravity, and the Gemini Enterprise Agent Platform.
For end users, the model powers the Rambler dictation system on Android Gboard in select regions and is enabled in the Gemini macOS application. Web field dictation support across Google Chrome is scheduled for rollout in an upcoming update.
