Google Debuts Gemini 3.5 Transcribe for Intelligent Voice Interaction

Google's new Gemini 3.5 Transcribe model provides high-precision speech-to-text with intelligent features that remove filler words and handle self-corrections. It offers a 70% latency improvement over previous models and supports more than 85 languages for global use. The model is now available for developers via API and is integrated into Android and macOS platforms to enhance voice-driven productivity.

Key Points

• Gemini 3.5 Transcribe features 'smart transcription' that automatically cleans up filler words and self-corrections for polished text.
• The model delivers high accuracy with a 2.6% Word Error Rate for pre-recorded audio and a 70% reduction in latency compared to its predecessor.
• It supports over 85 languages and includes advanced capabilities like multi-speaker attribution and function calling for complex workflows.
• Developers can access the model via the Live API for streaming or the Interactions API for recorded audio through Google AI Studio.
• The technology is being integrated across the Google ecosystem, including Gboard's Rambler feature and the Gemini app on macOS.

Sentiment

Cautiously optimistic but skeptical of marketing metrics

In Agreement

• The model offers top-tier accuracy, latency, and formatting for dictation tasks.
• The integration into Gboard and other consumer products is a significant step forward for mobile productivity.
• The model has the potential to significantly improve YouTube automatic captions and subtitle formatting for language learners.

Opposed

• Previous Google models like Chirp had severe hallucination issues in noisy environments that may persist.
• Word Error Rate (WER) is an insufficient metric because it doesn't capture errors in sentence structure and punctuation.
• The requirement for cloud-based compute is a drawback compared to local STT solutions.
• The rollout is confusing and restricted to very specific high-end hardware.