Google DeepMind has introduced Gemini 3.5 Transcribe, a dedicated speech recognition model engineered to turn raw audio straight into structured, agent-ready text without the usual pipeline scaffolding. Unlike legacy speech-to-text engines like Whisper that blindly transcribe every verbal stumble, the new model filters background noise, strips filler words, and handles live mid-sentence self-corrections on the fly.
According to benchmarks published by Artificial Analysis, Gemini 3.5 Transcribe posts a 4.0% Word Error Rate (WER) in streaming modes and 2.6% on batch non-streaming tasks, outperforming Chirp 3 while reliably parsing alphanumeric entities like order IDs across more than 85 languages.
Google is routing deployment through Google AI Studio and the Gemini Enterprise Agent Platform via two endpoints: gemini-3.5-transcribe-live for sub-second bidirectional streaming and gemini-3.5-transcribe for asynchronous batch audio featuring three-speaker diarization. The model also hooks into native clients like the Gemini macOS app to pipe formatted voice input straight into downstream tool calls.
By collapsing acoustic transcription, noise suppression, and disfluency cleanup into a single inference pass, the release kills the latency and compute tax of multi-stage voice middleware. Traditional STT providers now face an unappealing prospect: competing against foundation models that parse spoken intent rather than merely logging phonetic noise.