Google releases Gemini 3.5 Transcribe for real-time and pre-recorded speech-to-text with sub-4% WER

Gemini 3.5 Transcribe replaces earlier models with a single architecture that serves both live and batch use cases. The split between a Live API for streaming and an Interactions API for pre-recorded audio lets developers choose latency or accuracy without maintaining separate code paths. Speaker attribution and word-level timestamps are baked into the pre-recorded path, removing the need for third-party diarization libraries in call-log scenarios.

The claimed 4.0% WER for streaming and 2.6% WER for batch transcription are competitive benchmarks, but the material does not specify test conditions beyond “noisy, real-world environments.” Custom vocabulary support allows domain-specific terms to be recognized without fine-tuning, yet the material does not state whether this feature is available in both APIs or only in batch mode. Function calling is currently limited to the macOS app, so cross-platform voice agents must wait for wider rollout.

Latency improvements (70% faster time to final transcription) and support for 85 languages reduce the need for per-language models, but the material does not disclose pricing or rate limits. The model’s ability to handle live language switches and disfluency cleanup may simplify voice agent design, yet the experimental status of 3+ speaker diarization means multi-party meeting tools still require fallback logic. Vendor lock-in to Google’s API stack is the primary adoption cost.