Announcing Azure AI Speech LLM 2607: Better Multilingual Accuracy, Easier Customization

Azure AI Speech continues to deliver regular improvements to LLM Speech throughout the year, helping developers build more accurate and capable voice experiences. We’re announcing LLM Speech 2607, our latest model update, which brings improvements in multilingual recognition and mixed-language audio inputs, along with an enhanced experience for the phrase list feature.

Whether you’re building voice agents, meeting transcription solutions, contact center experiences, or online meeting applications, LLM Speech 2607 delivers transcripts you can trust for high-stakes, high-volume communication with no changes to your application architecture.

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

Key Advancements in Model Quality and Accuracy

This release delivers a decisive quality leap across Tier 1 and Tier 2 languages.

Building on the previous 2605 release, our internal evaluations also show that LLM Speech 2607 delivers measurable quality improvements across key scenarios:

• Mixed-language audio, where the language spoken changes mid-conversation, improving recognition when speakers switch languages during a conversation.
• Improved accuracy on Tier 2 and Tier 3 language locales, helping close the quality gap across a broader set of languages. In particular, capitalization, punctuation, and numeric and digit transcription are more robust.
• Domain-aware entity and person-name recognition, precisely capturing names, brands, and domain technology.
• Lower latency, improved up to 3x faster than the previous 2605 release. This is especially relevant for crisp, real-time responsiveness for interactive use.

Screenshot below shows a comparison of the previous 2605 release and the new 2607 release on two sample audio inputs. Punctuation and numeric transcription improvements are highlighted.

New Enhancement: Phrase Lists

A common challenge in speech applications is accurately recognizing domain-specific vocabulary such as product names, technical terminology, acronyms, company names, or industry jargon.

With LLM Speech 2607, we’re simplifying how a customer-provided phrase list can be enabled: use it as a dedicated parameter for providing important words and phrases instead of embedding custom vocabulary into a catch-all prompt. You can now specify recognition hints directly through a purpose-built field.

This makes customization simpler, cleaner, and easier to maintain in production applications. Phrase lists can include 2,000+ entities, providing a significant quality boost and allowing the LLM to be more context-aware. Further documentation can be found here.

Example request including phrase list

curl --location 'https://YourResourceName.cognitiveservices.azure.com/speechtotext/transcriptions:transcribe?api-version=2025-10-15' \
  --header 'Ocp-Apim-Subscription-Key: YourSpeechResourceKey' \
  --form 'audio=@"YourAudioFile"' \
  --form 'definition={
    "locales": ["en-US"],
    "phraseList": {
      "phrases": ["Contoso", "Jessie", "Rehaan"]
    }
  }'

Phrase Lists are especially useful for scenarios involving:

• Enterprise terminology
• Technical product names
• Industry-specific vocabulary
• Acronyms and abbreviations
• Customer-defined terms

Available through Fast API and Real-Time API

LLM Speech 2607 is available through the Fast API, enabling developers to bring the latest model improvements to both transcription and interactive voice experiences.

Get started

LLM Speech 2607 is the latest step in our ongoing investment to make speech AI more accurate, multilingual, and easier to customize. No customer action is required; this update will be deployed automatically on the service side. Try it out now in the Foundry Playground (Azure Speech – Speech to Text), or see REST API or SDK modalities in our public docs.