Azure AI Speech adds 25 languages and doubles processing speed
Microsoft has released LLM Speech 09-2026 for Azure AI Speech, adding 25 languages, including 11 Indic languages, and cutting processing time by up to 2x compared with the 2607 version. The update is available through the Fast Transcription API, according to the Microsoft Foundry blog.
- LLM Speech 09-2026 adds 25 languages to Azure AI Speech, including 11 Indic languages.
- Microsoft lists 50 languages covering 60+ locales in total.
- Processing is up to 2x faster than 2607, especially for 5-10 second audio at high concurrency.
- Internal evaluations show up to 25% better Token Error Rate and 37% better Entity Error Rate.
- No customer action is needed; the update is deployed automatically via the Fast Transcription API.
Wider language and locale coverage
The new model extends supported languages, with a focus on additional Indic-language coverage. The new Indic locales are Assamese, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and Urdu. Catalan, Malay and Welsh are also added. According to Microsoft, the full list now covers 50 languages and 60+ locales.
Microsoft says the model is meant for audio in which speakers use several languages in one input, as well as regional variants of a language, such as British English, American English and Indian English. As an example, the company describes a clinical consultation in Kuwait where the doctor speaks Arabic, the patient responds in Hindi, and notes are taken in English. Microsoft says LLM Speech handles this in a single transcription call.
Accuracy, speed and silent audio
Building on the 2607 release, Microsoft reports gains from internal evaluations: up to 25% improvement in Token Error Rate and 37% in Entity Error Rate. These figures come from internal and public benchmarks such as GigaSpeech and VoxPopuli. The company says the gains are particularly visible on earnings domain entities in financial scenarios.
On speed, LLM Speech 09-2026 delivers up to 2x faster processing than 2607. Microsoft says the improvement is most pronounced for short audio of 5 to 10 seconds at high concurrencies, meaning multiple streams at once. The release also continues work on silent and empty audio, so that silence is not treated as content. According to Microsoft, this reduces hallucinations and gives downstream applications cleaner output.
Multilingual mode and phrase lists
LLM Speech uses multilingual mode by default. Microsoft advises omitting locale hints when a file contains more than one supported language, since a locale hint is suited to genuinely single-language audio and is not a way to force recognition of a mixed-language conversation. If the locale of a file is known, specifying it can improve accuracy and minimize latency, according to the post.
The model can also detect the language automatically when no locale code is given. Extra options can be combined with fast transcription, including diarization, phraseList, locales, profanityFilterMode and channels. For product names, acronyms and domain terms, LLM Speech supports up to 2,000 phrases through the dedicated phraseList parameter.
Availability and what changes for users
Developers can try the model in the Foundry Playground (Azure Speech, Speech to Text) or through the REST API and SDK. Microsoft names voice agents, meeting transcription, contact center analytics and media captioning as target scenarios.
No customer action is required. Microsoft says the update will be deployed automatically on the service side, so existing users of the Fast Transcription API should receive the changes without migrating.
Microsoft Foundry: Announcing Azure AI Speech LLM 09-2026: Broader locale support
