Gemini 3.5 Transcribe release notes

Google DeepMind has introduced Gemini 3.5 Transcribe, a high-precision speech-to-text model designed to convert raw audio directly into polished, formatted text. The model improves upon previous iterations by handling background noise, complex jargon, and disfluencies—such as filler words and self-corrections—to produce clean transcripts suitable for real-time voice interactions.

Developer Access and API Implementations

Gemini 3.5 Transcribe is available through two distinct APIs to support different developer workflows:

Real-time streaming (gemini-3.5-transcribe-live): Provided via the Live API, this version delivers continuous, bidirectional streaming with sub-second latency for interactive voice applications.
Pre-recorded audio processing (gemini-3.5-transcribe): Provided via the Interactions API, this version is optimized for recorded audio, meetings, and call logs, featuring speaker attribution and word-level timestamps.

Developers can access these models through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.

Technical Capabilities and Performance

Gemini 3.5 Transcribe introduces several "smart transcription" features and performance improvements over the previous Chirp 3 model:

Transcription Precision and Intelligence

Disfluency Cleanup: The model automatically removes filler words (e.g., "ums" and "ahs") and handles self-corrections (e.g., "let’s meet Tuesday—no, Wednesday").
Custom Vocabulary: The model adapts to specialized jargon and unique spellings through provided custom vocabulary lists.
Multilingual Support: It automatically detects and transcribes over 85 languages, including various regional accents and dialects.
Multi-speaker Identification: For pre-recorded audio, the model attributes speech to up to three speakers with timestamps (support for more than three speakers is currently experimental).
Function Calling: In the Gemini macOS app, the model can delegate complex tasks, such as file analysis or image generation, to other Gemini models via function calls.

Benchmarks and Accuracy

As measured by Artificial Analysis, Gemini 3.5 Transcribe achieves an average Word Error Rate (WER) of 4.0% for streaming and 2.6% for non-streaming use-cases. On the FLEURS benchmark across top languages and locales, it achieves a 5.50% WER in streaming mode and 5.04% WER in non-streaming mode.

Compared to Chirp 3, the model shows a 70% improvement in time to final transcription.

Product Integrations

Gemini 3.5 Transcribe is integrated across several Google surfaces to enable context-aware voice interaction:

Gboard (Android): The "Rambler" feature uses the model to transform spoken thoughts into formatted text and allows voice-driven edits and style changes.
Google Antigravity: The model utilizes screen context and chat history to increase transcription accuracy for file names and active documents.
Gemini app (macOS): Enables natural speech transcription and voice commands that pair with screen context to summarize local files or generate images.
Google AI Studio: Available in Build mode for voice-driven application development ("vibe coding").
Chrome: Coming soon, the model will enable "talk to type" functionality in any web field.

Ecosystem Adoption

Several developer platforms have integrated the Gemini Live API to facilitate the deployment of voice-driven interfaces, including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents. Additionally, companies such as Vivo, Intellitek Health, and Lingopal have reported positive results regarding the model's latency and language support.