Google released Gemini 3.5 Transcribe, a speech-to-text model that handles 85 languages while automatically removing filler words and correcting verbal mistakes in real time. The model achieves a 4.0 percent word error rate in streaming mode, a 70 percent latency reduction compared to its predecessor Chirp 3.
The core capability here differs from standard transcription. Gemini 3.5 Transcribe doesn't just convert speech to text. It cleans up the output during transcription itself. Filler words like "um" and "uh" disappear. Self-corrections and grammatical stumbles get smoothed into proper written form. This happens as users speak, not after transcription completes.
Latency matters for real-time applications. A 70 percent reduction means faster response times for live captioning, voice assistants, and accessibility tools. The 4.0 percent word error rate puts it in line with professional-grade transcription systems, though exact comparisons depend on language and audio quality.
The function-calling feature extends the model's utility. Gemini 3.5 Transcribe can pass transcribed text to other Gemini models for downstream tasks. This enables workflows where speech input triggers additional processing. A user speaks instructions, the transcriber captures the text, and other models execute actions based on that input.
Language coverage at 85 options addresses global markets. Most transcription services focus on a handful of major languages. Broad language support reduces barriers for non-English speakers and enables localized applications in underserved regions.
Google positioned this as part of its Gemini family, suggesting integration with existing Gemini products and APIs. Developers can embed Gemini 3.5 Transcribe into applications through Google Cloud or integrate it with Gemini's broader ecosystem. The reduction in latency and error rates makes this practical for production use cases that previously required external transcription services.
Competitors in speech-to-text include OpenAI's Whisper, which also handles multiple languages, and traditional services like Amazon Transcribe or Microsoft Speech-to-Text. Gemini 3.5 Transcribe's real-time correction and cleaning differentiate it from basic transcription. Whisper, while open-source and accessible, doesn't perform the same level of on-the-fly cleanup.
Applications span customer service, accessibility, note-taking, and voice interfaces. Call centers benefit from automatic quality improvement in transcripts. Accessibility tools become more effective when background noise and stammering disappear from captions. Developers building voice-first applications gain a cleaner foundation for downstream processing.
The latency improvement opens new possibilities for live interactions. Faster transcription means users see text sooner in captioning applications. Voice assistants respond more quickly. Real-time translation pipelines built on top of transcription become more viable.
Google's emphasis on cleaning transcription output reflects a shift in how speech-to-text gets deployed. Raw accuracy matters less if the output requires heavy post-processing. A system that delivers clean, written text from spoken input reduces engineering work downstream. Developers spend less time building correction layers and more time on application logic.
The Chirp 3 predecessor already demonstrated Google's speech model capabilities. This update improves speed and usability without abandoning language coverage. It signals Google's ongoing investment in voice interfaces as computing input methods become less keyboard-dependent.