Microsoft released Azure AI Speech LLM 09-2026 on 26 October 2026, adding 25 languages to the Fast Transcription API and reducing processing time compared with the 2607 model. The update is service-managed, so customers do not select the new model or change clients. It is available through the Fast Transcription API.
Azure LLM Speech is the large language model used through enhanced mode on the shared fast transcription endpoint. Microsoft says it is used for transcription and translation, including captions, subtitles, meeting recordings, voicemails and contact centre transcripts. It is not called through a separate LLM Speech API.
Language coverage
The release extends Azure LLM Speech to 25 additional languages and brings the overall list to 50 languages across more than 60 locales. Microsoft names 11 new Indic locales: Assamese, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and Urdu. Catalan, Malay and Welsh are also listed among the additions.
The model runs in multilingual mode by default when the request omits locale hints. That matters for recordings where speakers change language inside the same file, such as a meeting or consultation that moves between Arabic, Hindi and English. If the audio is genuinely single-language, specifying a locale such as en-US can improve transcription quality and reduce latency.
Quality and latency changes
Microsoft says internal evaluations show up to a 25% improvement in Token Error Rate and up to a 37% improvement in Entity Error Rate. The measurements are described as covering internal and public benchmarks, including GigaSpeech and VoxPopuli. Microsoft highlights gains for earnings-domain entities in financial scenarios.
The release also cuts processing time by up to 2x compared with the 2607 model. Microsoft says the change is most visible on short audio of 5 to 10 seconds at high concurrency. That points to voice agents, call analysis and other systems where many short audio streams are processed in parallel.
Microsoft also says the model is less likely to treat silent or empty audio as content. For downstream systems that summarise, classify or route transcripts, that is a practical change. False text from silence can become input to later automation unless it is filtered before use.
API and configuration
Applications call the transcriptions endpoint with multipart form data and enable enhanced mode for the task. The task can be transcribe or translate. For translation, the request must include a target language code.
The same request can include diarisation, phrase lists, locale hints, profanity filtering and channel settings. Phrase lists are intended for product names, acronyms and domain terms, and the release supports up to 2,000 phrases through the phraseList parameter. Prompts can guide output style, including lexical formatting, but prompts unrelated to speech tasks are typically ignored.
The response includes combinedPhrases for the full text and phrases for segment-level and word-level detail. The response format is consistent with existing fast transcription and batch transcription outputs, but there are limits for translation. Word-level duration and offset fields are not supported for translation, diarisation is not supported for translation, and confidence is always returned as 0.
Prerequisites and limits
Developers need an Azure subscription and a Microsoft Foundry or Speech resource in a supported region. LLM Speech is available in the new Foundry portal, not the classic portal. Microsoft also provides REST and SDK options.
Audio files must be less than five hours long and smaller than 500 MB. Supported formats include WAV, MP3, OPUS/OGG, FLAC, WMA, AAC, ALAW in WAV, MULAW in WAV, AMR, WebM and SPEEX. For long audio files, Microsoft recommends upload from a public URL rather than inline audio.
The SDK material covers Python 3.9 or later, .NET 8.0 or later, Node.js LTS and Java Development Kit 8 or later. Authentication can use an API key or Microsoft Entra ID. Microsoft recommends keyless authentication with Microsoft Entra ID for production and requires the Cognitive Services User role for that route.
No customer action is required for the 09-2026 model update, and Microsoft did not announce a pricing change for this release. Teams that already use fast transcription should still review configuration if they depend on locale hints, phrase lists, translation output or retry behaviour under high concurrency.
Choosing where LLM Speech fits in an application, how its transcripts feed agents and which language or prompt settings are acceptable are Core AI questions at ECS, alongside agentic AI, Azure AI Foundry, the EU AI Act and responsible AI. Sessions and workshops give attendees room to compare product configuration with practitioner experience on Azure AI Foundry and governed AI outputs. Bring the speech workloads, language requirements and retry assumptions that are difficult to settle from documentation alone.
Photo: Coolcaesar, CC BY-SA 4.0, via Wikimedia Commons, cropped
This article was written by the ECS editorial team with the assistance of Amal Agent, run.events' AI assistant, which supported the research, drafting and proofreading.