Azure IoT guidance for DigiCert Global Root G1 trust-store updates in sovereign clouds
June 26, 2026Best practices for Infrastructure as Code CI/CD on Azure
June 26, 2026
Customer service runs on conversations that are hard to have and harder to practice. A frustrated customer with a billing charge that was never fixed. A cancellation that turns into a retention call. A confused caller who needs patience more than information. These are the moments that decide whether someone stays a customer, and they are exactly the moments a new agent has had the least chance to rehearse. As AI now absorbs routine contact volume, the hard conversations are increasingly the ones left for people, and Gartner found that 84% of customer service leaders plan to add new skills to the agent role to meet that shift.
The usual ways to build that skill all have a catch. Roleplay with a human coach works, but it is expensive, inconsistent, and almost impossible to schedule at scale. Shadowing live calls puts a trainee in front of a real customer before they are ready. Classroom training and scripts teach the words but not the pressure of a live, unscripted exchange. The one skill you most want to drill, staying calm and effective while a real person is upset, is the one you can least afford to practice on real people.
Live Voice Practice closes that gap. It is an open-source solution accelerator that turns Azure Voice Live into a real-time roleplay coach. A trainee picks a scenario, has an actual spoken conversation with an AI customer, and the moment the call ends, gets scored against a rubric with specific, evidence-backed feedback. The hard call becomes something you can rehearse as many times as you need: safely, consistently, and with a number at the end that tells you whether it landed.
Azure Voice Live is the real-time, speech-to-speech service underneath it. It lets a person hold a spoken conversation with an AI model that listens, understands, and answers out loud with low latency, so the exchange feels like a phone call rather than typing into a chatbot.
Why Hard Calls Resist Training
Training someone for a difficult call is not really a content problem. Most agents can recite the policy. The problem is doing the right thing under pressure, in real time, while a customer pushes back, and the standard training tools are weak exactly where it counts.
The standard tools fall short in four related ways. Human roleplay is inconsistent and subjective: two coaches play the same angry customer differently and grade the result on gut feel, so a score is really an impression and “good” drifts from one trainee to the next. It also does not scale, because a coach can run only a handful of sessions a day while a contact center may be onboarding hundreds of agents who each need repeated reps on every hard scenario.
The deepest problem is that the most valuable rehearsal is the one you cannot ethically stage. You do not learn to defuse an angry customer by practicing on a real one, so the highest-stakes skill ends up getting the least deliberate practice. Live Voice Practice is built to lift all four limits at once: a consistent AI customer, an objective rubric, unlimited repeatable sessions, and no real customer at risk.
A Concrete Example: The Frustrated Billing Customer
To make this real, take one scenario that ships as an example: a frustrated customer with an unresolved billing charge.
The setup is simple and familiar. The customer sees a charge on their account that does not make sense. They already contacted support once, nothing was fixed, and now they are calling back annoyed. The agent’s job is not just to find the answer. It is to absorb the frustration without escalating it, gather the right account details, explain clearly what happened, resolve it, and leave the customer with proof that it is handled.
That is a lot to do well, and all of it has to happen live. In a classroom you would talk about empathy and ownership. Here the agent has to actually open with it, while a customer who is already irritated keeps interrupting. This is the kind of conversation Live Voice Practice is designed to rehearse, and it is the scenario the demo below walks through end to end.
Inside a Practice Session
A session is built to feel like a real call, and each part of that experience maps to a specific Azure Voice Live capability worth understanding.
It opens on a set of practice scenarios. The agent picks the frustrated billing customer and chooses a Live Voice Agent to play them. That agent appears as an avatar, a face and a voice on screen, because Voice Live can drive a lip-synced video avatar in real time rather than returning audio alone. The effect is not cosmetic: practicing against a face you have to read restores the social pressure of a real difficult call.
Then the conversation runs, and it is a genuine back-and-forth rather than a turn-by-turn script. The microphone stays open while the AI customer talks, so the agent can interrupt and be interrupted. That barge-in, speaking and listening at the same time with low latency, is the feature that separates a real call from a walkie-talkie, and it is what lets the customer cut in the moment the agent stalls. Voice Live also handles noise suppression and echo cancellation, so a trainee can practice from an ordinary desk instead of a sound booth.
The AI customer stays grounded in the exact scenario the agent chose. When it needs those details, Voice Live invokes a function the app exposes, in the middle of the conversation, and gets the scenario back. That is function calling doing something concrete: anchoring the customer in the selected situation instead of a generic complaint.
Every turn lands in a live transcript as it is spoken. When the agent ends the call and asks for the analysis, that transcript, not a recording someone has to re-listen to, is the evidence the scoring runs on. After a short pause the result appears: an overall score out of five and a clear pass-or-fail result, with an explanation of how it was calculated alongside it.
The detail comes in three layers. First, the score is broken down by rubric criterion, with clarity, empathy, information gathering, completeness, and closure each scored and justified against what happened on the call. Next, the weakest criteria are surfaced first, and for each one the report explains what it measures, why the call scored that way, the evidence from the transcript, and a concrete next step. Finally, a plain-English summary lays out strengths alongside gaps. Seconds after hanging up, the agent knows what they did well, what cost them points, and what to drill next.
See it in action: a single live call with the frustrated billing customer runs end to end, from picking the scenario to a Pass score, a criterion-by-criterion breakdown across clarity, empathy, information gathering, completeness, and closure, and a set of recommended next steps for the agent’s next practice run.
What Makes the Feedback Worth Trusting
For a score to be useful, two things have to hold: it has to mean the same thing on every call, and the agent has to be able to see why they earned it. Live Voice Practice is built for both consistency and transparency.
The rubric is the contract: a set of named, weighted criteria that defines what a good call looks like for a given scenario. Because the same rubric runs on every session, “good” stays consistent across agents and attempts, the one thing human roleplay can never guarantee. And every criterion returns a reason, not just a number, naming what the agent did, the moment in the transcript, and the next step, so a low score teaches instead of just judging.
Two analyses run on every call, side by side. One evaluates the conversation: an Azure OpenAI model scores the transcript against the rubric and returns structured results. The other assesses pronunciation using Azure Speech, so clarity of delivery is measured, not just the words. Both run when the agent asks for the analysis, and the result comes back as a single report.
The practice itself is grounded in reality. Scenarios are derived from real training materials, and the AI customer’s behavior is shaped by curated example transcripts, so the agent on the other end of the line behaves the way real customers do rather than the way a model imagines they might.
The result is feedback that is consistent because the rubric never changes between calls, specific because it is tied to evidence from the conversation, and immediate because the agent sees it the second they hang up. The same evidence is there for a coach to review later, which turns a private practice run into something a team can actually learn from.
The Azure Services Doing the Work
Strip away the interface and Live Voice Practice is a thin orchestration over four Azure services, with one of them doing most of the heavy lifting.
Azure Voice Live is the engine. It is a single, real-time speech-to-speech API that folds together what used to take three separate services: speech-to-text to hear the trainee, a generative model to decide what the AI customer says, and text-to-speech to say it out loud. The Voice Live API is now generally available, so the accelerator stands on a current, supported foundation rather than a preview that might shift under it.
At Build 2026, Microsoft sharpened almost every capability this Solution Accelerator depends on, and all of them live inside the one API. The interruption handling behind natural barge-in, the lip-synced avatar, built-in noise suppression and echo cancellation, and the function calling the app uses to keep the customer grounded in the chosen scenario are all native to Voice Live. Microsoft also introduced a new generation of all-in-one speech-to-speech models, including GPT-Realtime 1.5 and the Azure-Realtime model, that make the exchange faster and more natural across languages, and refreshed the voice and avatar stack with new Neural HD V3 voices and full-body avatars. Using one service instead of stitching these pieces together is why Live Voice Practice can hold a lifelike conversation without building a custom audio pipeline.
The scoring runs on Azure OpenAI, reached through an Azure AI Foundry endpoint. When the agent asks for the analysis, a conversation analyzer sends the transcript and the scenario’s rubric to a chat model and asks for a structured, criterion-by-criterion result. That step is what turns a free-form conversation into a comparable score with written justifications.
Azure AI Speech adds the second opinion. A pronunciation assessor scores how clearly the agent actually spoke, so delivery is measured alongside content. These two analyses run concurrently on the same Analyze request and return as one combined report, which is why a thorough evaluation comes back in seconds rather than one stage after the other.
Azure Cosmos DB holds the scenarios and their rubrics. Because they live as data rather than code, a team adapts the tool to its own business by editing documents: point the scenarios at real training material and the rubrics at the criteria it actually grades on, with no redeploy.
Live Voice Practice simplified architecture
Built for the Enterprise
Training data and call transcripts can be sensitive, so Live Voice Practice was built with secure deployment in mind from the start, not bolted on later.
It offers two topologies. The Basic deployment uses public endpoints and is meant for fast evaluation and demos. The Network-isolated deployment is built for production on the Azure AI Landing Zone, Microsoft’s reference architecture for running AI workloads under a Zero Trust posture. Every Azure service sits behind a private endpoint, users on the public internet reach the app through an Azure Application Gateway with a Web Application Firewall, outbound traffic is forced through Azure Firewall, and administrators who need direct access can optionally reach a jumpbox through Azure Bastion. The same accelerator that lets you try the idea in an afternoon also gives you a hardened path, on a blueprint your platform team likely already recognizes, to run it for real.
Live Voice Practice network-isolated reference architecture on the Azure AI Landing Zone
In that topology the app runs in an Azure Container App, and every dependency (Azure AI Foundry, AI Speech, AI Search, Cosmos DB, Key Vault, Storage, and App Configuration) is reached through a private endpoint inside the virtual network. Public users come in through the Application Gateway and its WAF, and all outbound traffic leaves through Azure Firewall. When administrators need direct access to the network, they can optionally reach a jumpbox through Azure Bastion.
Microsoft Entra ID sign-in adds the layer that makes this a coaching tool rather than a single-player demo. Once agents sign in, each one keeps a history of their past practices, and trainers get a view across everyone, so individual reps become a coachable cohort whose progress you can actually see.
Deploy It Yourself
Getting it running is a single azd up. The command provisions the Azure resources, deploys the app, and hands you a URL. From there you pick a scenario, choose a voice agent, and run your first call. To make it yours, point the scenarios at your own training materials and set the rubric to the criteria your team actually grades on.
From One Call to a Coaching Loop
The real payoff is not a single strong rehearsal. It is what happens when every agent on the team can run that rehearsal on demand.
Picture an entire team rehearsing the hard conversations on demand: the same scenario, the same fair rubric, instant feedback after every attempt, with no roleplay partner to schedule and no real customer at risk. Because the feedback is scored, “they are getting better” stops being a hunch and becomes something you can track, agent by agent and criterion by criterion, over time.
The hardest part of customer service was never knowing the policy. It is staying calm, clear, and human when a real person is upset, and that is a skill you build by doing it, safely, again and again. Live Voice Practice is how you let people do exactly that. Pick a scenario and run your first call.