Google has introduced Gemini 3.5 Transcribe, a state-of-the-art speech-to-text model designed for both real-time voice applications and recorded audio processing. Unlike a single unified solution, the model is delivered through two distinct endpoints, each optimized for different use cases:
gemini-3.5-transcribe: Handles pre-recorded audio files via the Interactions API, providing high-accuracy offline transcription.gemini-3.5-transcribe-live: Supports bidirectional streaming through the Live API, enabling real-time voice interfaces with low latency.
According to evaluations by Artificial Analysis, the model achieves an average word error rate (WER) of 4.0% in streaming mode and 2.6% in non-streaming mode. This marks a significant 70% improvement in time-to-final-transcription compared to Chirp 3, Google's previous generation model.
The system automatically detects and transcribes more than 85 languages, including seamless mid-sentence code-switching, making it a strong contender for multilingual applications.
Deployment Considerations
Gemini 3.5 Transcribe is available exclusively as a managed API service. There are no open weights or self-hosted deployment options. This is a strategic decision by Google to maintain control over the technology and ensure consistent performance, but it means that organizations must integrate with Google's infrastructure.
Key Differences Between Endpoints
Developers should plan carefully around the split between the two endpoints, as they do not share identical feature sets, usage limits, or pricing structures. The streaming endpoint prioritizes low latency, while the offline endpoint focuses on accuracy. Understanding these trade-offs is essential for optimizing cost and performance in production environments.
via MarkTechPost