Microsoft Launches Faster Azure AI Speech LLM 2607

Microsoft has rolled out LLM Speech 2607, the latest iteration of its Azure AI Speech model. Building on the previous 2605 release, this update introduces substantial performance upgrades, most notably reducing latency by up to three times. This speed boost is designed to enable more responsive, real-time interactions for voice agents, contact centers, and live meeting transcription services without requiring developers to modify their existing application architecture.

The new model brings significant quality enhancements to multilingual and mixed-language environments. It is better equipped to handle conversations where speakers switch languages mid-sentence, and it improves transcription accuracy for Tier 2 and Tier 3 language locales. Specifically, the update refines how the system handles capitalization, punctuation, and numeric data, while also improving domain-aware recognition for brand names, technical terms, and individual people.

For developers needing to customize vocabulary, Microsoft has overhauled the phrase list feature. Instead of forcing users to embed custom terms into a general prompt, LLM Speech 2607 introduces a dedicated parameter field for recognition hints. Developers can now feed more than 2,000 distinct entities directly into this field, making it easier to maintain custom vocabularies like product names or industry jargon in production.

The LLM Speech 2607 model is available immediately through the Fast API and Real-Time API, and it can be tested in the Azure Speech-to-Text section of the Foundry Playground. Because Microsoft is deploying the update automatically on the service side, existing customers do not need to take any manual action to begin using the new model.