📰 Key Summary

This news comes from the Google DeepMind blog, not internal data, translated directly:

Google has launched a new speech-to-text model, Gemini 3.5 Transcribe, built for accurate and intelligent speech recognition. It’s already powering the Gemini App, Android Rambler, and the macOS Gemini App, and developers can now access it through the Gemini API (Google AI Studio) and the Gemini Enterprise Agent Platform. The model ships as two APIs: a real-time streaming version, gemini-3.5-transcribe-live, which delivers sub-second-latency, bidirectional streaming via the Live API — great for voice assistants and live captions; and a pre-recorded audio version, gemini-3.5-transcribe, which processes recordings like meetings and calls through the Interactions API, with speaker labeling and word-level timestamps. Key features include smart semantic correction (for example, automatically fixing “let’s meet Tuesday — no wait, Wednesday”), automatic removal of filler words like “um” and “uh” with clean formatting; function calling support, so tasks like image generation or file analysis can be handed off to other Gemini models (currently limited to the macOS App); custom vocabulary support, so it can adapt to specialized terminology and unusual spellings; support for over 85 languages, with automatic detection and handling of different accents and dialects; and for pre-recorded audio, speaker labeling for up to three speakers with timestamps (four or more speakers is experimental). According to testing by Artificial Analysis, word error rate (WER) averages 4.0% in streaming mode and 2.6% in non-streaming mode, and the model stays accurate on alphanumeric content like zip codes and order numbers even in noisy environments. Compared to the previous-generation Chirp 3 model, the new model shows significant improvements in both accuracy and latency.

💬 JudyAI Lab Take

Looking at how Gemini 3.5 Transcribe is designed, what stands out about this release is that speech recognition has evolved from plain transcription into something that understands semantic context and self-corrects.

For AI builders, this points to a clear industry trend: the competitive edge in voice interfaces is shifting from “hearing accurately” to “understanding correctly.” Things like semantic-level correction (“let’s meet Tuesday — no wait, Wednesday”) and automatic filler-word removal with clean formatting show that the model isn’t just an acoustic recognizer anymore — it’s folding language understanding directly into the transcription pipeline. At the same time, splitting real-time streaming and pre-recorded audio into two separate APIs (sub-second latency vs. word-level timestamps and speaker labeling) shows that a single model has to serve two very different use cases — voice assistants and meeting transcripts — each with its own tradeoffs between latency and accuracy. The fact that word error rate stays stable on alphanumeric content (zip codes, order numbers) even in noisy environments also points to where the real recognition challenge lies in practice: structured information, not casual conversation.

If you want to try it hands-on, start by testing the custom vocabulary feature in the Gemini API and see how well it picks up the specialized terms in your own domain.

📅 Original Source Info

🔗 Further Reading