Azure LLM Speech adds 25 languages, 11 of them Indic, and runs up to 2x faster

Microsoft has added 25 languages to its Azure LLM Speech transcription model, 11 of them Indic, according to a Foundry developer blog post published on October 26. The company also claims up to 2x faster processing than the previous release.

The release, called LLM Speech 09-2026, brings the service to 50 languages covering more than 60 locales. It is reached through Azure’s Fast Transcription API, and the post says no customer action is required because the update is deployed automatically on the service side.

Which languages arrive and how fast it runs

The new Indic locales are Assamese, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and Urdu. Catalan, Malay and Welsh complete the 25 new languages.

Microsoft frames the update around mixed-language audio. Its example is a clinical consultation in Kuwait where the doctor speaks Arabic, the patient answers in Hindi and the notes come out in English, all handled in a single transcription call. The model also deals with regional variants of one language, such as British, American and Indian English.

On quality, the post cites internal evaluations showing up to a 25% improvement in Token Error Rate and 37% in Entity Error Rate. These were measured across internal data and public benchmarks, including GigaSpeech and VoxPopuli, with the largest gains on earnings-related entities in financial scenarios. Microsoft also says the model now treats silent or empty audio less often as speech, which should reduce hallucinated text.

The speed gain is not uniform. Microsoft says the 2x figure applies particularly to short clips of 5 to 10 seconds when many streams run at once, which is the pattern of a voice agent or a contact center. Longer files may see less benefit, and the post gives no figures for them.

How to use it today

LLM Speech runs in multilingual mode by default. Microsoft advises leaving out locale hints when a file mixes several supported languages, and using a hint only for audio that is genuinely in one language. The sample REST call in the post, which uses API version 2025-10-15, combines options such as diarization (separating speakers), a profanity filter mode, a locale list and a phrase list. The phrase list accepts up to 2,000 terms for product names, acronyms and domain vocabulary.

The model can be tried in the Foundry Playground under Azure Speech – Speech to Text, or through the REST API and SDKs. The post does not name regions, and it does not say whether the older model remains selectable after the automatic update.

A pricing question the post leaves open

The angle here is cost. The announcement gives no price, but Microsoft’s Azure Speech pricing page carries a footnote saying LLM Speech currently shares its price and SKU ID with Fast Transcription. In practice, the added languages and the speed gain do not appear to come with a separate charge, at least for now. The page text we retrieved showed placeholders instead of dollar amounts, so the actual hourly rate has to be checked in the Azure pricing calculator.

The same pricing page lists Microsoft’s own MAI-Transcribe-2 model as a separate line, with a limited-time promotional offer running until December 31, 2026. That suggests Microsoft is selling more than one transcription option inside Azure, and buyers will need to compare them on language coverage and price rather than assume LLM Speech is the only route.

The release also follows quickly on its predecessor. LLM Speech 2607, announced in a September 10 post, focused on multilingual accuracy and a dedicated phrase list field. The new version therefore arrives about six weeks later, which points to a steady update cadence rather than a one-off launch.

What the source does not establish

All accuracy and speed numbers come from Microsoft and are described as internal evaluations alongside public benchmarks. The post does not publish per-language error rates, and it does not say how the newly added languages score compared with the existing ones. Readers building products for Tamil, Telugu or Welsh speakers will want to test their own audio before relying on the aggregate figures. The 25% and 37% gains are also stated as upper bounds (“up to”), not averages.

The update lands in the same period as other Foundry changes. Our earlier report on the September Foundry update covered the model and agent side of the platform, while this release extends the speech layer that voice agents depend on. For teams already using the Fast Transcription API, the practical change is that more of their callers can now be transcribed in one workflow, with no migration step announced.