Live Voice Practice Platform

A real-time AI-powered voice training platform for call-center and customer-service practice using Azure Voice Live and Azure AI services.

Overview

This repository implements a voice-based practice and training platform for call-center agents, using real-time AI-driven customer simulations. The application acts as a virtual practice partner, allowing trainees to engage in realistic conversations with an AI-powered customer and receive structured feedback after the session.

Key capabilities include:

• Real-time voice conversations with an AI-powered customer
• Configurable training scenarios derived from training materials
• Customer behavior guided by curated example transcripts
• Automatic conversation transcript capture
• Post-session scoring and structured feedback
• Evaluation using reusable rubrics
• Secure deployment in a private cloud environment

In a typical session, a trainee selects a practice scenario and begins a real-time voice conversation with an AI-driven customer through Azure Voice Live, while transcripts and interaction events are captured. After the session ends, the interaction is evaluated asynchronously to generate structured feedback. Scenarios, rubric-based evaluation, and example transcripts enable more structured and repeatable training workflows.

Architecture

The diagram above shows the network-isolated (ZTA) topology: the browser reaches the app through Application Gateway WAF v2, the Container App runs the React frontend + Flask/WebSocket backend, and all data-plane traffic to Azure AI Foundry / Voice Live, Speech, Cosmos DB, AI Search, Key Vault, App Configuration and Storage flows through private endpoints inside the VNet. Outbound traffic from the jumpbox subnet is forced through Azure Firewall. For a deeper dive into the runtime flow see docs/how-it-works.md.

Note: the Azure AI Speech service (used by the app for the Live Voice avatar) is also deployed and reached via a private endpoint, but is not drawn in the diagram. It belongs with the GenAI App Backing Services group.

Deployment

This project supports two deployment modes:

• Basic (public endpoints) — fastest path, suitable for dev/demo:

  azd up

• Network-isolated (private endpoints + jumpbox + Azure Firewall + Application Gateway WAF v2) — production-grade:

  azd env set NETWORK_ISOLATION true
  azd provision   # from your workstation
  # then connect to the jumpbox via Bastion and run azd deploy + postProvision there

  For public access in network-isolated mode, opt in with azd env set PUBLIC_INGRESS_ENABLED true before azd provision, then use the public ingress runbook to configure domain/cert and validate with a real microphone. If you only need private access (jumpbox, Bastion, Azure Virtual Desktop in the spoke, or ExpressRoute/VPN), leave PUBLIC_INGRESS_ENABLED unset — the Application Gateway is not deployed and the runbook is not needed.

Defaults at a glance

Parameter	Always	Only when NETWORK_ISOLATION=true
Container App identity (useUAI=false)	System-assigned (default)	(same)
ACS media egress firewall rules (enableAcsMediaEgress)	n/a (no firewall)	enabled (default) — opens UDP 3478-3481 / TCP 443+3478-3481 to AzureCloud for Speech avatar / ACS Calling / Teams Media
Application Gateway WAF v2 skeleton (publicIngress.enabled)	n/a (Container App ingress is public)	disabled (default) — opt in with azd env set PUBLIC_INGRESS_ENABLED true only when you need a public entry point; otherwise reach the app from the jumpbox / Bastion, an Azure Virtual Desktop in the spoke, or an ExpressRoute/VPN into the VNet

Override any of these with azd env set (e.g. azd env set USE_UAI true, azd env set ENABLE_ACS_MEDIA_EGRESS false, azd env set PUBLIC_INGRESS_ENABLED true) before azd provision. The public-ingress runbook only applies when PUBLIC_INGRESS_ENABLED=true; with the default the Application Gateway WAF v2 is not deployed and the runbook is not needed.

Deployment guides

• Full deployment guide — prerequisites, environment variables, both deployment modes, validation, and teardown.
• Network isolation quick reference — workstation/jumpbox split, subnets, firewall allow-list, and troubleshooting.
• Public ingress runbook — BYO domain/certificate setup and real microphone validation.

Optional user sign-in

The app works without user login for practice and analysis. Configure Microsoft Entra ID sign-in after deployment when you want:

• Past Practices for each signed-in user
• Trainer views such as All Practices
• user identity from Azure Container Apps built-in authentication

Guide: Configure user authentication with Microsoft Entra ID.

Speech is provisioned by the AILZ Bicep template (deploySpeechService=true by default, AILZ v1.1.4+). To opt out:

azd env set DEPLOY_SPEECH_SERVICE false

Local Development

1. Clone the repository

2. Copy the environment template

cp .env.template .env

3. Configure the required Azure credentials

4. Build and run the backend

./scripts/build.sh
cd backend
python src/app.py

Open the application at:

http://localhost:8000

Project Origin and Acknowledgments

This project builds upon the sample Voice Live API Sales Coach. This repository extends the original implementation to support a voice-based practice platform for customer-service and call-center training scenarios, introducing configurable scenarios based on training materials, rubric-driven evaluation, and curated example transcripts.

The foundational architecture and part of the initial implementation originate from the original repository and its contributors. We thank the authors for making their work available as open source.

License

This project is licensed under the MIT License.

See the LICENSE file for details.