Google Launches Gemini 3.5 Transcribe with Smart Speech Cleanup and Sub-Second Streaming

Google has officially introduced Gemini 3.5 Transcribe, its next-generation speech-to-text foundation model tailored for real-time voice agents, interactive applications, and speech analytics. Succeeding Chirp 3, the model is exposed through two distinct operational endpoints: gemini-3.5-transcribe via the Interactions API for batch audio processing, and gemini-3.5-transcribe-live via the Live API over WebSockets for continuous, bidirectional streaming with sub-second latency. The model supports automatic language detection across 85+ locales, speaker diarization for up to three speakers, word-level timestamps, and custom vocabulary biasing. Crucially, it introduces a dedicated smart transcription mode that removes filler words, formats output, and automatically resolves real-time self-corrections into clean text. For AI practitioners building conversational interfaces, automatic speech recognition (ASR) has long presented a painful architectural bottleneck. Traditional speech engines transcribe audio verbatim, forwarding conversational stutters, filler words, and mid-sentence corrections directly into downstream large language models. This inflates context window token usage, introduces parsing ambiguity, and adds turn-taking latency. By embedding semantic disfluency filtering directly into the transcription layer while achieving Word Error Rates of 2.6% on recorded files and 4.0% on live streams—alongside a 70% reduction in time to final transcription relative to Chirp 3—the model significantly streamlines the voice-to-action path for virtual agents and customer support systems. This release illustrates an accelerating industry trend: the consolidation of speech recognition and conversational intelligence at the ingestion boundary. Instead of deploying complex, cascaded microservices—such as separate audio noise gates, acoustic decoders, regex cleaners, and downstream prompt orchestrators—cloud platforms are embedding contextual understanding directly into specialized multimodal models. This approach mirrors the broader movement toward unified realtime voice systems where the transcription layer is natively aware of intent, tooling hooks, and conversational dynamics. In practice, engineering teams should evaluate their voice architectures to capitalize on the separation between raw transcription and intent generation. Voice bot developers can configure smart streaming mode to eliminate bespoke transcript-cleaning logic in their orchestration layer. However, teams handling compliance-mandated recordings, such as legal or clinical intake, must ensure workloads explicitly use verbatim mode to prevent the model from normalizing spoken corrections. Finally, infrastructure engineers must design around strict API boundaries, including WebSocket session limits and multi-speaker concurrency constraints, when provisioning voice runtime services.