Gemini-3.5-Transcribe
Points and comments are a snapshot, not live.
Google releases Gemini 3.5 Transcribe, a speech-to-text model with 2.6% WER.
Google introduced Gemini 3.5 Transcribe, a speech-to-text model for real-time and pre-recorded audio. It achieves 2.6% WER on non-streaming and 4.0% on streaming use cases per Artificial Analysis, a 70% latency improvement over Chirp 3. Features include smart transcription (removing filler words, handling self-corrections), custom vocabulary support, multi-speaker ID for up to three speakers, and over 85 languages. Available via the Gemini API in Google AI Studio and Enterprise Agent Platform, it also powers Rambler on Gboard Android and the Gemini app on macOS.
What commenters are saying
Commenters are cautiously interested but raise several issues. A top complaint is that smart transcription removes intended words (e.g., 'I hesitated to check it' becomes 'I should have verified'), though the model offers both Smart and Verbatim modes. Some users report hallucination problems with previous models like Chirp, but early testers say 3.5 Transcribe does not hallucinate on silence or noise. Others note the feature is limited to Pixel 10+ and newer Samsung phones. Several recommend alternatives: Voxtral Mini 3b for multilingual industry jargon, and local models like Whisper for timestamp accuracy paired with Gemini for correction.