Gemini 3.5 Transcribe Replaces Legacy ASR with Native Disfluency Handling

Google unveiled Gemini 3.5 Transcribe, a dedicated speech-to-text model built directly on Gemini's multimodal audio understanding architecture. Released in public preview through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform, the model ships with two dedicated endpoints: a synchronous endpoint (gemini-3.5-transcribe) for processing recorded audio and a bidirectional streaming endpoint (gemini-3.5-transcribe-live) engineered for sub-second live interactions. The model succeeds Chirp 3, achieving a 2.6% Word Error Rate (WER) in non-streaming workloads and 4.0% in streaming mode, while supporting automatic language detection across more than 85 locales, multi-speaker diarization, word-level timestamps, and custom vocabulary biasing. Speech-to-text architectures have historically operated as rigid acoustic decoders, dumping verbatim transcriptions riddled with filler words, stuttering, and spoken corrections directly into downstream large language models. This design inflates LLM prompt token costs and introduces latency, forcing engineering teams to deploy dedicated pre-processing layers or write complex prompt instructions to clean conversational artifacts. Gemini 3.5 Transcribe addresses this by introducing native Smart Transcription, which semantically resolves on-the-fly corrections (such as updating dates spoken in error), strips disfluencies, and outputs properly formatted syntax without extra parsing steps. This release reflects a broader paradigm shift across cloud and AI infrastructure toward task-optimized distillations of frontier multimodal models. Rather than relying on isolated, legacy Automatic Speech Recognition (ASR) engines or using expensive, general-purpose LLMs for basic audio parsing, cloud providers are deploying lightweight, specialized audio endpoints. By leveraging the semantic foundations of the broader Gemini family, this approach retains deep language comprehension while minimizing compute cost and latency for voice agent architectures. For platform engineers and AI architects, adopting Gemini 3.5 Transcribe requires evaluating the trade-offs between smart and verbatim modes. Smart mode provides significant latency and token-saving benefits for agentic voice UIs and meeting summarization pipelines, but compliance, legal, and audit-heavy workflows will still require explicit verbatim flags to preserve raw acoustic fidelity. Furthermore, teams integrating the Live API over WebSockets should assess network resilience and client connection handling to maintain consistent sub-second streaming across variable network conditions.