Google announces Gemini 3.5 Transcribe for AI-powered speech-to-text (arstechnica.com)

Google has announced Gemini 3.5 Transcribe, a new AI model focused on enhancing speech-to-text capabilities by refining voice input. Designed to eliminate filler words like "ums" and self-corrections, this model is set to improve user experience across Google’s ecosystem, having already powered the Gboard "Rambler" feature on the Pixel 11. Notably, Gemini 3.5 Transcribe claims to be 70% faster than its predecessor, Chirp 3, with a reduced live-speech error rate of 5.5%, a significant advancement given that the previous model stood at 7.32%. The significance of this rollout lies in its potential to streamline communication and improve transcription accuracy in diverse settings, accommodating up to three speakers in recorded audio across 85 languages. Beyond mere recognition of spoken words, the model's ability to interpret intent and contextualize jargon reflects a leap forward in natural language processing for AI. However, the reliance on AI to filter and rephrase responses raises concerns about fidelity in specific contexts, making it essential for users to consider when and how to deploy this functionality effectively.