This article answers commonly asked questions about the Voice Live API. If you can't find answers to your questions here, check out other support options.
General
What scenarios does Voice Live support?
Voice Live API supports a wide range of real-time, natural voice interaction scenarios: contact centers, automotive assistants, accessibility applications, virtual tutors and learning companions, multilingual public service agents, HR support, and training. Used by customers like eClinicalWorks and the Government of Malta.
How does Voice Live compare to the Azure OpenAI Realtime API? When should I choose which?
Voice Live API enhances the AOAI Realtime API by offering: expanded model selection (including GPT-Realtime, GPT-5, GPT-4.1, PHI), more natural voice options, more supported speech languages, avatar integration, advanced semantic voice activity detection (VAD), seamless Microsoft Foundry Agent Service integration, and telephony integration via Azure Communication Services.
Which regions does Voice Live support?
Voice Live is available in 10+ Azure regions. For more information, see Region support.
What is the tokens-per-minute threshold?
The current limit is 100,000 tokens per minute per resource. Customers can request an increase. For more information, see Speech service quotas and limits.
Generative AI Models
What generative AI models are supported?
Voice Live supports OpenAI models in Microsoft Foundry, Phi-based LLMs, and SLMs. For more information, see Voice Live overview. Voice Live also provides an option to bring-your-own model (preview).
How do I choose the LLM for my use case?
Consider: accuracy (Azure Speech-based models are more robust for noisy audio), existing LLM solutions (reuse prompts and grounding data), latency (text-based LLMs can have slightly higher latency), inference cost (smaller models can be more cost-effective).
What is response instruction?
Response instruction guides model behavior and context. Define agent personality, specify questions, control response formatting. Responses should be concise and normalized for optimal audio synthesis.
What is response temperature?
Response temperature controls randomness of output. Lower values = deterministic, higher = creative. Adjust temperature or Top-P, not both.
Speech Input
Which languages does Voice Live support?
Voice live supports 146 languages/locales for input, 151 for output, 600+ neural voices. See Voice Live language support.
How do I get the live transcripts from the call?
Use text output events. Details at Voice live API reference.
What is a phrase list?
Domain-specific terms to improve recognition. Limit to <500 words/phrases. See How to customize Voice Live.
Are there other ways to improve speech input recognition accuracy?
Use Azure AI Custom Speech models. Configure multiple custom models per language. See How to customize Voice Live.
Speech Output
What voices does Voice Live support?
Native audio output with preferred model, Azure Speech in Foundry Tools TTS voices (600+ voices, 150+ locales, 30+ Neural HD voices). Custom voice models via Professional Voice Fine-tuning. For more information, see Voice Live API supported languages.
How do I pick a voice?
Use Voice Gallery in Microsoft Foundry Speech Playground. Consider gender, age, capability, style, personality.
What is voice temperature?
Controls expressiveness. Higher = dynamic/emotive, lower = neutral. Applies to Neural HD voices.
What is speaking rate?
Controls agent's speech speed.
What is a custom lexicon?
Define pronunciation rules for specific words. See How to customize Voice Live.
What is Custom Voice?
Create brand-specific synthetic voices using your own audio data. See How to customize Voice Live.
What is Avatar support?
Pair speech output with visual avatars for multimodal experiences.
What is Custom Avatar?
Photorealistic digital human using Azure AI TTS. Built from video recordings, tailored to specific actor's appearance and voice.
Conversational Enhancements
What is the difference between Azure Semantic VAD and Basic Server VAD?
Azure Semantic VAD is more noise robust and accurate for detecting utterance boundaries.
What is EOU (End of Utterance) detection?
Uses context to determine if a user finished speaking or just paused.
How does noise suppression work?
Filters background noise based on advanced technology.
How does echo cancellation work?
Removes echo of agent’s own voice picked up by microphone.
Function Calling
Does Voice Live support function calling?
Yes, including asynchronous function calling.
Is there model context protocol (MCP) support?
Currently, MCP is supported in with model mode with the exception of phi models. Further with Foundry (new) agents. It's not supported with Foundry (classic) agents.
Pricing
Where is the pricing listed?
How do I estimate the cost based on my use case?
Estimate by audio minutes; tokens are billing unit. See pricing and token usage and cost estimation.
Are there separate quota and throttling limits for voice-live?
Yes, quota applies specifically to Voice Live API (default: 100k tokens/min).
Additional
Does this service provide an SDK?
Yes, SDKs for Python, C#, Java (Preview) and JavaScript/TypeScript (Preview) are available. See Voice Live - Reference - Voice Live SDK.
Does this service include content filtering?
Yes, content filtering is included.
Can you modify or disable the content filtering in Voice Live API?
No. If you need custom content filtering, you can use the bring-your-own-model feature.
Is SIP supported?
SIP is currently not supported.
Next steps
- Learn more about How to use the Voice Live API
- See the Voice Live API reference
- What's new