Azure AI Speech adds locale coverage and keeps mixed-language transcription automatic

Microsoft’s [October 26 announcement](https://devblogs.microsoft.com/foundry/announcing‑azure‑ai‑speech‑llm‑09‑2026/) introduces LLM Speech 09‑2026 through the Azure AI Speech Fast Transcription API. The service update expands its language coverage, with a practical configuration choice for recordings that switch between languages. The notice names 11 new Indic locales: Assamese, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and Urdu. It also names Catalan, Malay and Welsh, and points to a full support list covering 50 languages and 60‑plus locales. Those named additions are not the entire language inventory. Developers should check the complete support list for the specific speech they need to transcribe.

The [documented default](https://devblogs.microsoft.com/foundry/announcing‑azure‑ai‑speech‑llm‑09‑2026/) is multilingual transcription. Microsoft tells developers to omit locale hints when a file contains more than one supported language. The service detects and selects the appropriate language from the audio without a locale code. A locale hint is intended for genuinely single‑language audio. It is not a way to force recognition of a conversation that moves between languages. That makes the contents of the recording, rather than a customer’s account language or application interface, the relevant basis for choosing this setting.

For product names, acronyms and domain terms, the [announcement describes](https://devblogs.microsoft.com/foundry/announcing‑azure‑ai‑speech‑llm‑09‑2026/) a dedicated `phraseList` parameter supporting up to 2,000 phrases. That is a phrase‑list limit; the notice does not supply a general prompt‑token budget. Phrase hints and locale hints should not be treated as interchangeable controls. The sample calls the `speechtotext/transcriptions:transcribe` operation with API version `2025‑10‑15` and a `multipart/form‑data` request containing an audio file and a definition. Its `enhancedMode` configuration enables the feature, sets the task to `transcribe` and includes a prompt for lexical‑format output. The notice also lists diarization, phrase lists, locales, profanity filtering and channels as configuration options that can be combined with fast transcription. These are documented fields, not a request OpenTools has executed. The announcement’s full curl example has locale values without JSON string quotes, so it should be checked against the linked request‑configuration documentation before being copied into an application.

Microsoft says the update is [available through the Fast Transcription API](https://devblogs.microsoft.com/foundry/announcing‑azure‑ai‑speech‑llm‑09‑2026/), with REST, SDK and Foundry Playground routes. It says no customer action is required for the model update, which will be deployed automatically on the service side. That does not verify the rollout state of an individual Azure resource. This report covers language support and request behavior. It does not reproduce Microsoft’s error‑rate or processing‑speed comparisons. OpenTools has not tested audio, verified language‑by‑language quality, or established regional availability and pricing. Applications still need to review transcripts against their own recordings and requirements before relying on them. *Microsoft Visitor Center Building 92, photographed May 29, 2022. Archive company‑context image; not Azure AI Speech infrastructure or the October 2026 rollout. Photo: [Lectrician1](https://commons.wikimedia.org/wiki/User:Lectrician1), via [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Microsoft_sign_at_the_Microsoft_Visitor_Center_Building_92.jpg); [CC BY‑SA 4.0 International](https://creativecommons.org/licenses/by‑sa/4.0/deed.en). Resized and cropped for display; image adaptations retain the same licence.*