Microsoft has expanded Azure AI Speech with 25 additional languages while cutting transcription time for some workloads. The LLM Speech 09-2026 model now supports 50 languages across more than 60 regional variants through the Fast Transcription API.
The additions include 11 Indic languages: Assamese, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and Urdu. Catalan, Malay and Welsh are also new. Multilingual mode can identify language changes within one recording, while developers can still provide locale hints for single-language audio.
Microsoft says internal and public benchmark testing found improvements of up to 25 percent in token error rate and 37 percent in entity error rate, particularly for financial terminology. Those are vendor-reported maximum gains rather than a guarantee for every recording. The new model also processes short, highly concurrent audio workloads up to twice as fast as its predecessor and is less likely to treat silence as speech.
Developers can use phrase lists containing as many as 2,000 names, acronyms or domain terms. The service also supports speaker separation and channel settings. Existing applications may receive the model as the default is updated, so teams with accuracy-sensitive workflows should test their own accents, noise conditions and terminology before relying on the reported gains.