Transcribe 25 new languages at up to 2x the speed. LLM Speech 09-2026 expands Azure AI Speech to 25 additional languages—including 11 Indic languages—while cutting processing time. Whether you’re building voice agents, captioning systems, or contact center analytics, your multilingual workflows just got faster and more accurate, building on the multilingual and phrase-list improvements introduced in LLM Speech 2607.

Picture a clinical consultation in Kuwait where the doctor speaks Arabic, the patient responds in Hindi, and notes are taken in English—LLM Speech handles it all in a single transcription call.

Azure LLM Speech service, accessed through the Fast API, delivers accurate transcripts faster and with necessary (model will be updated to the default), enhancing voice agents, meeting transcription, contact center analytics, or media captioning scenarios.

Try it out now in the Foundry Playground(Azure Speech – Speech to Text), or see REST API and SDK options in our public docs.

Advancements with this release

** Broader language coverage**

LLM Speech 09-2026 model extends supported languages, including additional Indic-language coverage, so more conversations can enter the same transcription workflow. This supports customer scenarios where speakers speak not only multiple languages in one audio input, but also regional variants of a language (i.e. British English, American English, and Indian English). For example, our model now better handles a clinical transcript in Kuwait where the doctor and patients are speaking a mix of Arabic, Hindi, and Indian English.

New Indic locales: Assamese, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu, Urdu. Additional new languages include: Catalan, Malay, and Welsh. See language and locale support for the full list of 50 languages covering 60+ locales.

** Improved recognition quality**

Building on 2607, internal evaluations show measurable accuracy gains, up to 25% improvement in Token Error Rate and 37% in Entity Error Rate metric across internal and public benchmarks like GigaSpeech and VoxPopuli, in this 09-2026 release, particularly on earnings domain entities for financial scenarios.

** Faster processing**

LLM Speech 09-2026 delivers up to 2x faster processing than 2607, returning transcripts sooner for high-volume and time-sensitive workloads, particularly for short audio (5-10 second length) at high concurrencies (multiple streams at once).

** Robustness on silent and empty audio**

Continued quality work helps ensure silence isn’t treated as content, reducing hallucinations so downstream applications get cleaner output.

Multilingual transcription, built for real conversations

Many conversation scenarios follow multiple language settings. A customer may begin in English, switch to Spanish to explain a detail, and return to English. LLM Speech uses multilingual mode by default: omit locale hints when a file contains more than one supported language. A locale hint is useful for genuinely single-language audio, not a way to force recognition of a mixed-language conversation.

Call the REST API with cURL

curl --location 'https://YourResourceName.cognitiveservices.azure.com/speechtotext/transcriptions:transcribe?api-version=2025-10-15'
--header 'Content-Type: multipart/form-data'
--header 'Ocp-Apim-Subscription-Key: <YourSpeechResourceKey>'
--form 'audio=@"YourAudioFile.wav"'
--form 'definition={
  "enhancedMode": {
    "enabled": true,
    "task": "transcribe",
    "prompt": ["Output must be in lexical format."]
  },
  "diarization": {
    "maxSpeakers": 2,
    "enabled": true
  },
  "phraseList":{
    "phrases":["Kenichi Kumatani","John McDonough"]
  },
  "locales":[en-in, gu-in, en-us],
  "profanityFilterMode": "Masked"
}'

In your POST request to the transcriptions endpoint, use the multipart/form-data content type with the audio file and the request body properties.

The following example shows how to transcribe an audio file with a specified locale. If you know the locale of the audio file, you can specify it to improve transcription accuracy and minimize the latency.

• Replace YourSpeechResoureKey with your Speech resource key.
• Replace YourResourceName with your Speech resource name.
• Replace YourAudioFile with the path to your audio file.

You can transcribe audio in the input language without specifying a locale code. The model automatically detects and selects the appropriate language based on the audio content.

You can combine extra configuration options with fast transcription to enable enhanced features, such as diarization, phraseList, locales, profanityFilterMode, and channels.

Tip: Pair with phrase lists

For product names, acronyms, and domain terms, add a phrase list to your request. LLM Speech supports up to 2,000 phrases through the dedicated phraseList parameter. See how to use phrase lists with LLM Speech in our docs.

Availability

LLM Speech 09-2026 is available through the Fast Transcription API.

Get started

The LLM Speech 09-2026 release is the latest step in our ongoing investment to make speech AI more accurate, multilingual, and efficient. No customer action is required; this update will be deployed automatically on the service side.

Try it out now by REST API, SDK, or in the Foundry Playground (Azure Speech – Speech to Text). Explore our docs:

• LLM Speech overview
• Use the fast transcription API
• API migration and request configuration

The post Announcing Azure AI Speech LLM 09-2026: Broader locale support appeared first on Microsoft Foundry Blog.