Microsoft has updated Azure AI Speech to better transcribe conversations that move between languages, a common failure point for voice agents, meetings and contact centers. LLM Speech 2607 also targets more accurate capitalization, punctuation, numbers, personal names and specialist terms across a wider range of language locales.
Microsoft says internal evaluations show latency improving by as much as threefold compared with the previous 2605 release. That could make the change especially noticeable in interactive voice applications, where even a short delay disrupts a conversation. The performance figures are Microsoft’s own and can vary with the audio and deployment.
Customization is also becoming more explicit. Developers can place important names, brands, acronyms and technical terms in a dedicated phrase-list field instead of mixing them into a general prompt. A list can hold more than 2,000 entities, which should make production configurations easier to maintain.
LLM Speech 2607 is available through Azure’s Fast API and Real-Time API, as well as the Foundry Playground, REST API and SDK routes. Microsoft says the service-side rollout is automatic, so existing customers do not need to change their application architecture to receive the model update.