Azure AI Fundamentals: One-Hour Scenario-Based Prep for AI-900 2024 Candidates
A brisk, one—hour study guide for Microsoft Exam AI—900 that teaches you to recognize AI workloads and map them to the right Azure services—Azure Machine Learning, AI Vision, AI Language/Speech, and Bot Service—favoring prebuilt capabilities and practical, no—code patterns. It pairs clear definitions, workplace scenarios, pitfalls and quiz—checks with responsible—AI principles, an exam—day checklist, and a time map so you can choose the simplest, safest solution under pressure.
By MyAudioBooks.ai ·
Listen free: Azure AI Fundamentals: One-Hour Scenario-Based Prep for AI-900 2024 Candidates
Astori Publishing Presents: Azure AI Fundamentals: One-Hour Scenario-Based Prep for AI-900 2024 Candidates Prologue Imagine this: It's a crisp morning in Seattle, 2024, and you're seated in a bustling tech conference room, laptop open, as a Microsoft engineer demos a simple Azure AI model that spots defects on a factory assembly line in real time. The screen flickers with imagesaflawless parts gliding by, then one with a tiny crack, flagged instantly by bounding boxes drawn by an invisible hand. The room buzzes with murmurs: "No code needed?" someone asks. The engineer nods, explaining how prebuilt Azure services turn raw data into decisions, but warns of pitfalls like biased training sets that could skew results for night-shift footage. You lean forward, notebook in hand, realizing this isn't just tech wizardryait's the gateway to certifying your grasp of AI fundamentals, where one exam could unlock doors in cloud computing. This scene unfolds in the heart of today's AI landscape, anchored in Microsoft's Azure ecosystem, where the AI-900 certificationaoften called Azure AI Fundamentalsaserves as your entry point. Launched to bridge beginners to the world of artificial intelligence, it demands no coding prowess but a sharp eye for mapping real-world problems to the right tools. Picture the key players: Azure Machine Learning, the organizer that structures your data experiments; Azure AI Vision, the eagle-eyed scout for images and documents; Azure AI Language, the insightful analyst of text and sentiment; and Azure Bot Service, the conversational guide that orchestrates chats. These aren't abstract entities; they're the heroes in stories like the retailer's demand forecast or the call center's sentiment triage, drawn from Microsoft's 2024 exam objectives and service guides that emphasize practical pattern recognition over deep dives. At its core, the AI-900 exam probes a tantalizing paradox: In an era where machines learn from data faster than everathink deep learning's breakthrough in 2012, when neural networks crushed image recognition contestsahow do you choose the simplest, most responsible path amid hype and hazards? Artificial intelligence, or AI, encompasses systems mimicking human smarts, while machine learning lets models evolve from patterns without rigid rules. Yet urgency mounts: With AI powering everything from fraud detection to chatbots, missteps like overfitting (where a model memorizes training data but flops on new inputs) or ignoring privacy could derail deployments. Training providers stress this tensionaopt for prebuilt capabilities first, but always mind ethical guardrails like fairness and transparency. As you absorb these foundations, envision yourself not just passing the exam but wielding AI with confidence, spotting when a scenario calls for supervised learning's labeled predictions or unsupervised clustering's hidden insights. This guide promises to equip you with that clarity: grouped vocab, workplace vignettes, quiz-checks, and traps to sidestep, all in a brisk one-hour flow. Pause here, on the edge of discoverya what if the next chapter transforms how you see AI's potential and pitfalls? Dive in, and let's map the path together. End of Prologue Per Microsoftas 2024 exam objectives and service documentation, AI-900 is a fundamentals certification that emphasizes concepts and Azure service capabilities rather than code, with a passing score commonly set at seven hundred on a oneathousandapoint scale. The objectives focus on recognizing appropriate AI workloads, understanding basic machine learning tasks, and selecting the right Azure services. Training providers echo this framing and advise you to map scenarios to services, prefer prebuilt capabilities first, and watch for common pitfalls. You will hear that guidance once here, and then I keep attribution light so the pacing stays brisk and useful. You are listening to a focused, oneahour guide designed to help you think like the exam. You will work with grouped vocabulary, short workplace scenarios, common traps, and quick quizachecks. You do not need to write code to pass. You do need to recognize tasks, metrics, and service names. Brief, regular practice in Azure AI Studio or service portals helps the names stick. The mindset is pattern recognition: map the scenario, match the service, mind the constraints. We start with machine learning basics and computer vision, and we prime your ear for the language and patterns you will reuse when we later turn to language and conversation. According to the 2024 Azure Machine Learning product guides that major training providers teach from, your machine learning goals for this exam are modest and precise: name the main learning approaches, choose sensible metrics for classification and regression, and pick Azure features that solve a scenario without overcomplicating it. Start with the learning types. Artificial intelligence is the umbrella for systems that perform tasks associated with human intelligence. Machine learning is a subset where models learn patterns from data rather than relying on hardcoded rules. Deep learning uses multilayer neural networks to capture complex patterns in images, audio, and language. A quick historical anchor helps: in 2012, a deep neural network won the ImageNet vision contest by a large margin, a result that pushed deep learning into the mainstream for vision tasks. Within learning types, supervised learning trains on labeled examples to predict known outcomes; classification predicts categories, while regression predicts numeric values. Unsupervised learning discovers structure without labels, such as grouping customers by behavior. Reinforcement learning trains an agent to make sequential decisions with feedback signals; for AIa900, you recognize it conceptually. Lock in the data terms. A feature is an input variable the model uses to learn. A label is the target you want the model to predict. A training set is the data used to fit the model. A validation set tunes choices without peeking at final results. A test set stays held out to estimate performance on unseen data. If analogies help, training and validation are rehearsals; the test set is opening night. Match metrics to tasks. For classification, accuracy is the share of correct predictions among all predictions. Precision is the share of predicted positives that are actually positive. Recall is the share of actual positives the model found. The F1 score balances precision and recall and is useful when classes are imbalanced. For regression, mean absolute error reports the average absolute difference between predictions and actuals, while root mean squared error emphasizes larger mistakes by squaring them before averaging and then taking the square root. Area under the receiver operating characteristic curve, often called AUCaROC, compares classifiers by how well they rank positives above negatives across decision thresholds. Think of mean absolute error as your average miss distance on a dartboard, while root mean squared error punishes the big misses that matter most. Map these ideas to Azureas workflow. An Azure Machine Learning workspace organizes assets like data, experiments, and models. Compute targets provide machines to train at scale. Environments capture package versions for reproducibility. Pipelines chain steps for preparation, training, and evaluation. Endpoints expose models for realatime or batch predictions. Automated ML tries many model and preprocessing combinations within your constraints to produce strong baselines without writing code. The Responsible AI dashboard in Azure Machine Learning helps you probe model behavior with interpretability, error analysis, and fairness checks. In practice, teams often start with Automated ML to benchmark several algorithms quickly, then use the dashboard to see where errors cluster before moving to deployment. Ground this with two scenarios. A retailer asks you to forecast weekly demand for each product and store. That is supervised regression. You engineer features from past sales, promotions, and holidays, train in Azure Machine Learning, use Automated ML to establish a strong baseline, and deploy the best model to a managed online endpoint for scoring. Now consider a fraud screening pipeline for card transactions. That is supervised classification with imbalanced data, so accuracy alone misleads. You emphasize precision and recall, consider the F1 score, and may use class weighting or resampling to handle rarity. This is general information, not individualized financial advice; production systems require regulatory review, human oversight, and appropriate controls. Expect a few traps. Overfitting shows up when a model shines on training data but stumbles on new data; simplify the model, regularize to penalize unnecessary complexity, or gather more representative data. Data leakage occurs when information from the future or the label slips into training features; split by time for timeabased problems and avoid computed features that peek ahead. Choosing the wrong metric is common; imbalanced problems need more than accuracy. Automated ML is powerful, but it still needs clean, correctly labeled data and a clear problem type. Getting splits and metrics right is half the battle on exam day. Quizacheck. Question: Which split best estimates final performance on unseen dataatraining, validation, or test? Pause. Answer: The test set. Explanation: The test set is held back from fitting and tuning, so it gives the least biased estimate. Quizacheck. Question: In an imbalanced classification task, which single metric is a safer first look than accuracyarecall or F1? Pause. Answer: F1. Explanation: F1 balances precision and recall when positives are rare. Quizacheck. Question: You want a noacode way to try many models and preprocessing options. Which Azure capability fits? Pause. Answer: Automated ML in Azure Machine Learning. Explanation: It automates model and feature processing within constraints you set. Per the 2024 Azure AI Vision, Custom Vision, and Azure Document Intelligence guides echoed across established training materials, your computer vision goals are to distinguish the main task types, map them to the right Azure services, and know when prebuilt beats custom. Vision is where deep learningas rise first registered with the public; the ImageNet breakthroughs made object recognition feel practical almost overnight. Set the task vocabulary first. Image classification assigns a single label to the whole image, like adefectivea or ano defect.a Object detection finds and localizes multiple items with bounding boxes, such as athree helmets visible.a Semantic segmentation labels each pixel to outline shapes with high precision. Optical character recognition, or OCR, extracts text from images. Azure Document Intelligence extends OCR by extracting structured fields, tables, and layout from documents like receipts and invoices. Face detection locates faces and landmarks; features that infer attributes or identify individuals carry stricter controls and regional limits and are treated cautiously on the exam. A quick realaworld note: postal services have relied on OCR for years to read addresses at scale, a reminder that prebuilt text extraction is often enough without custom training. Align the services to the tasks. Azure AI Vision provides image tagging, captions, OCR, and other image analysis features that work out of the box. Custom Vision lets you upload and label your images and then train classification or detection models tailored to your domain; you can host these models in Azure or export them to edge devices for onasite processing. Azure Document Intelligence targets forms and semiastructured documents, extracting keyavalue pairs and tables into structured output. Put simply, Vision is your prebuilt toolkit, Custom Vision is your tailored model shop, and Document Intelligence is your forms specialist. Apply this on a factory line. Cameras capture each finished part. If you only need a passaorafail outcome, image classification fits. If you must point to the defectas location, choose object detection. If the team needs a precise outline to measure a crack or scratch, semantic segmentation is the best match. In Azure, you start in Custom Vision with labeled examples and publish an endpoint for realatime inspection. Variants of this approach show up in electronics and automotive plants to flag missing components and surface flaws before products move downstream. Shift to backaoffice operations. A logistics team processes waybills and receipts at scale. OCR can read raw text from scans, but when the requirement mentions amounts, dates, vendor names, and tables, Azure Document Intelligence is designed for structured extraction. On exam wording that mentions keyavalue pairs or line items, prefer Document Intelligence over general OCR. When bandwidth is tight on the shop floor or trucks upload intermittently, exporting a Custom Vision model to the edge can reduce latency and keep the line moving. Avoid familiar traps. Mixing up classification and detection is common; classification labels the whole image, while detection finds and counts items. Reaching for a custom model too early wastes time; prebuilt image analysis or document models may already solve the need. Privacy signals matter; facearelated workloads require consent and policy compliance. A quick memory aid helps: classification is like tagging a photo, detection is like drawing boxes on a map to show where things are. Quizacheck. Question: The requirement is to pull totals, dates, and vendor names from scanned invoices into a database. Which capability should you chooseageneral OCR, Custom Vision, or Document Intelligence? Pause. Answer: Document Intelligence. Explanation: It extracts structured fields and tables beyond plain text. Quizacheck. Question: You must count hard hats in a site photo. Which task fits betteraclassification or object detection? Pause. Answer: Object detection. Explanation: You need to identify and locate multiple items, not just label the entire image. Quizacheck. Question: The prompt mentions drawing a precise outline around a paint defect. Which task is that? Pause. Answer: Semantic segmentation. Explanation: Segmentation labels pixels to produce detailed shapes. You now have the bedrock for this certification: you recognize learning types and data splits, you pick metrics that fit the task, and you map vision needs to the right Azure capability. Building on this foundation, you are ready to move from images to language and interaction, where you will connect text, speech, and conversational flows next. You now move from images to language and interaction. The goal here is practical: recognize text, speech, and conversation workloads, then map each one to Azure AI Language, Azure AI Speech, and Azure Bot Service with confidence. Per Microsoftas 2024 Azure AI Language and Speech documentation and the widely taught AIa900 curricula, prebuilt capabilities cover most exam scenarios; the task is to match the requirement to the right feature and note privacy or accuracy limits when input is sensitive or noisy. Think of a language pipeline as a relay raceaeach stage hands a cleaner baton to the next, from raw input to structured insight. Start by locking in the text analytics vocabulary you will hear. Naturalalanguage processing, or NLP, analyzes or generates human language. Sentiment analysis estimates whether text is positive, neutral, or negative. Key phrase extraction pulls out the main ideas. Named entity recognition finds and labels defined types such as people, organizations, locations, dates, and amounts. Personally identifiable information, or PII, detection flags sensitive items like names, emails, phone numbers, and identification numbers so they can be masked or redacted. Language detection identifies the primary language when it is not specified. Summarization condenses long passages while preserving the important points. Question answering retrieves or extracts direct answers from provided sources such as documents or knowledge articles; it replaces earlier QnA Makerastyle flows. A useful distinction helps here: entities are like the clearly printed fields on a formadates and amountsawhile key phrases are the headline takeaways you would jot in the margin. Place the Azure services that support these tasks. Azure AI Language hosts sentiment, key phrases, entities, PII detection, language detection, summarization, and question answering under one umbrella. Translator handles machine translation across many languages and can feed translated text into Language features when you want consistent downstream analysis. Azure AI Speech covers audio. Speech to text converts spoken audio into text. Text to speech synthesizes naturalasounding audio from text. Speech translation turns spoken words from one language into another. Speaker diarization separates who spoke when so later analysis can attribute each utterance to the right person. The same documentation and training materials emphasize a simple but durable pattern: use Speech for transcripts, then apply Language for insights. Work this into a realistic scenario. A support operations team wants to triage tickets automatically. You detect the language, run Translator if needed, then use Azure AI Language to extract key phrases and entities and to score sentiment so urgent cases surface when tone is negative. If the requirement mentions protecting customer privacy, you add PII detection to mask sensitive details before storage. The bestafit answer on the exam is a short, orderly pipeline: detect, translate, analyze. Now shift to audio. A contact center records calls and wants insights. Speech to text produces transcripts; if there are multiple speakers, you enable diarization so each utterance maps to the right person. Azure AI Language then extracts key phrases, entities, and sentiment. If leaders ask for quick summaries, you add summarization to produce call notes. This aSpeech first, Language seconda sequence appears frequently in both the documentation and training walkthroughs. Watch the common pitfalls. Sarcasm and cultureaspecific expressions can trip sentiment models, so highastakes decisions still need human review. Confusing key phrases with entities is another trap; key phrases are topicalike summaries, while entities are typed items such as dates or organizations. Translation can unify multilingual input but may change nuance; the conservative workflow is to translate for consistency and acknowledge small tradeaoffs in tone. If you store text or audio that includes PII, detecting and redacting sensitive fields before broader analysis is the safe default. When privacy and accuracy pull in different directions, the exam tends to reward answers that mention redaction first, then analysis. Quizacheck. Question: The requirement says aflag customer names and account numbers for masking before analysis.a Which feature do you apply firstakey phrases, sentiment, or PII detection? Pause. Answer: PII detection. Explanation: It identifies sensitive items for masking so later steps operate on protected text. Quizacheck. Question: You have a twoaspeaker audio file and need peraspeaker sentiment. What do you enable in the speech step before sending text to Language? Pause. Answer: Speaker diarization. Explanation: Diarization separates who spoke when so downstream analysis is attributed correctly. Quizacheck. Question: You must analyze posts from several languages in a common scoring space. Do you translate first or score natively? Pause. Answer: Translate first with Translator, then apply Language features. Explanation: This yields consistent scoring across languages, while noting small nuance shifts. With text and speech in hand, turn to conversational AI and dialog design. Your objectives stay practical: explain intents and entities, pair Conversational Language Understanding with question answering, and place Azure Bot Service and Bot Framework Composer in the flow. Per the 2024 Azure Bot Service materials and the updated Azure AI Language guides, modern bots combine intent recognition, knowledgeabased answers, and routing across multiple skills. Older namesaLanguage Understanding Intelligent Service, known as LUIS, and QnA Makerahave been superseded by Conversational Language Understanding, or CLU, and the question answering features in Azure AI Language; the guides recommend using the updated terms. Set the core conversation terms clearly. A bot is an automated agent that interacts over chat or voice. An intent represents a useras goal, such as atrack ordera or areset password.a An utterance is an example phrase that teaches the model what a user might say. An entity is a data element inside an utterance, like an order number, city, or date. A dialog is the planned flow of prompts, confirmations, and actions the bot follows to complete a task. Orchestration routes a request to the best capability when multiple skills are available. If you like images, intents are the destination on a transit map, entities are the tickets with the details, and the dialog is the route you take to get there. Map the tools to these needs. Azure Bot Service hosts the bot and connects it to channels such as web chat and Microsoft Teams. Bot Framework Composer provides a visual way to author dialogs, prompts, and language generation so you can test flows quickly. CLU recognizes intents and entities. The question answering feature within Azure AI Language indexes documents and returns short answers, serving as the successor to earlier knowledge bases. Together, these pieces let you mix singleaturn answers with multiaturn tasks. In practice, teams often lean on question answering to handle FAQ volume and reserve CLUadriven dialogs for actions that require collecting detailsaan approach the documentation and training examples reinforce. Apply this to an FAQ scenario. A shipping company wants a bot that answers questions from an existing handbook and can escalate to an agent. You create a bot in Azure Bot Service, add a question answering project with the handbook as a source, connect the chosen channel, and include a human handoff path. If the requirement mentions followaup questions on the same topic, you keep the dialog minimal and let question answering handle coverage from the indexed content. This keeps setup light and preserves the handbook as the single source of truth. Now consider a taskaoriented scenario. A travel assistant must handle abook flight,a achange booking,a and arefund,a capturing dates and cities. You define intents for each goal and entities for dates, airports, and confirmation codes in CLU. In Composer, you build dialogs that prompt for missing details, confirm what the user said, and support corrections. If the scenario adds multiple skillsasay, hotel and car rentalayou enable orchestration so the bot routes each request to the correct skill. Think of orchestration as the dispatcher, directing each message down the right path without the user needing to know what is under the hood. Avoid frequent traps on the exam. Mixing up question answering and intent recognition is common; FAQs that live in documents fit question answering, while multiastep tasks need intents, entities, and dialogs. Omitting a fallback is another pitfall; a resilient bot always includes a graceful aI didnat get thata path with logging for improvement. When you see legacy names, choose the updated ones: CLU instead of LUIS, and question answering instead of QnA Maker. For voice channels, remember that Speech features capture and synthesize audio, while the same understanding stack processes the text. The simplest capable solution usually winsaprefer prebuilt question answering before custom orchestration unless the scenario demands task flows. Quizacheck. Question: The requirement says ause the company handbook as the single source of truth and return short answers with minimal setup.a Which capability do you chooseaCLU or question answering? Pause. Answer: Question answering. Explanation: It indexes documents and returns concise answers without designing intents or dialogs. Quizacheck. Question: A bot must handle abook hotel,a aextend stay,a and acancel,a and extract dates and locations. What two constructs do you define in the understanding model? Pause. Answer: Intents and entities. Explanation: Intents capture the goals; entities capture the details like dates and cities. Quizacheck. Question: Your bot supports both FAQs and a atrack ordera workflow. How does it decide where to send each user message? Pause. Answer: Orchestration. Explanation: Orchestration routes messages to the right skill, such as question answering for FAQs and CLU plus dialog for tasks. You now have a single mental model for language experiences: Azure AI Language turns content into structure, Azure AI Speech links audio and text, and Azure Bot Service coordinates the interaction. Next, you wrap these capabilities with responsible AI so your choices are not just effective but also fair, safe, private, transparent, inclusive, and accountable. Per Microsoftas responsible AI guidance updated in 2024aan approach echoed in mainstream training for this examaresponsible use centers on six principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. You finish by weaving these principles through the workloads you have already learned, then close with a practical review, an examaday checklist, a oneahour time map, and a brief producer note so the narration lands smoothly. Start with plain definitions you can recognize in scenarios. Fairness means similar users receive similar outcomes; you check error rates by group and correct gaps you find. Think of a reliable bathroom scale: it should read the same weight no matter who steps on it. Reliability and safety mean the system behaves as intended in normal conditions and fails safely in edge cases; you test, monitor, and add guardrails so behavior degrades gracefully instead of abruptly. Privacy and security mean you minimize data, protect it in storage, transit, and use, and restrict access. Inclusiveness means you design and test for diverse users and contexts, including accessibility needs. Transparency means you document what the system does and why in language people can understand. Accountability means specific people own decisions, with human oversight and a path for appeals. Connect those principles to tools you already know. In Azure Machine Learning, the Responsible AI dashboard brings interpretability, error analysis, and fairness assessment into one place so you can see where errors concentrate before deployment. You can log predictions and outcomes to watch for data driftainput patterns that shift over timeaand to track performance. For content risks, Azure AI services include safety filters that reduce harmful outputs. For privacy and security, you pair encryption, access controls, private networking, and data minimization. For transparency, you maintain concise model cards that summarize purpose, data, and limits, along with data descriptions and change logs. At fundamentals level, clarity beats code release; the guidance emphasizes documentation users can act on rather than internal source code. A quick reality check ties these ideas to the public record. Independent audits of early face analysis systems documented higher error rates for some skinatone and gender combinations, turning fairness from an abstract concept into a concrete testing requirement. Meanwhile, widely covered incidents of chat systems producing confident but incorrect statements show why reliability and safety matter; fluent output is not the same as trustworthy output. The guidance responds by emphasizing evaluation, documentation, and human oversight for higherarisk decisions. Work through a fairness caselet in the style the exam favors. A loan preascreening model rejects more qualified applicants from one demographic than others. You segment errors by group, confirm the disparity, and retrain with more representative data or adjust decision thresholds, then reaevaluate with fairness metrics. This is general information, not individualized financial advice; financial deployments require qualified professional review, organizational policy, and regulatory compliance. Reliability and safety appear differently on a factory camera system that flags missing hard hats. You stressatest with glare, shadows, and varied gear, set action limits for uncertain cases, and monitor the false alarm rate so the system remains useful rather than noisy. Privacy and security arise when you store call transcripts for sentiment analysis. You first detect and redact personally identifiable information, then apply language analysis, and you restrict access to the original audio. Avoid the pitfalls that recur across domains. Declaring abiasafree dataa is a red flag; the pragmatic posture is assess, mitigate, and monitor. Treating transparency as arelease the codea misses the mark at this level; clear documentation, user guidance, and contact points for redress matter more. Overatrusting a strong validation score without checking subgroup performance or temporal drift risks failure after launch; plan monitoring and regular reaevaluation windows. And for safety controls, do not rely on filters alone; pair them with policy, training, and escalation paths. A useful habit is to pause and ask, where would this system most likely fail, and who would be affected? Use a twoalens view for fastamoving capabilities. In claim world, model providers describe rapid gains in accuracy and utility for large language and vision models, often citing internal benchmarks and striking demos. In the public record through mida2024, thirdaparty tests show strong results on some datasets but also hallucinations, brittleness under distribution shift, and uneven performance across languages and demographics. Open questions include how to quantify and reduce those failures at scale, how to communicate limits to end users without overwhelming them, and how to set thresholds for when human review is mandatory. What to watch next are standardized evaluations, stronger contentasafety tooling, and governance frameworks such as impact assessmentsaterms that are starting to appear in examalevel discussions. Quizacheck. Question: A model performs well on average but shows much higher false negatives for nightashift footage. Which principle is implicated first, and what is a practical first step? Pause. Answer: Fairness; analyze errors by context and retrain with more representative nightashift data, then reatest with fairness metrics. Quizacheck. Question: A customer asks you to store full audio and text for broad reuse ajust in case.a Which principle pushes back, and what is the safer default? Pause. Answer: Privacy and security; minimize data, redact PII, and apply strict access and retention controls. Quizacheck. Question: A stakeholder wants to remove safety filters to boost recall. Which principle is at risk, and what do you propose? Pause. Answer: Reliability and safety; keep filters, tune thresholds within policy, and add human review for uncertain cases. Here is what to watch in the near term so your studying stays current without chasing headlines. Expect clearer exam phrasing around humanainathealoop boundariesahuman oversight at key stepsamore references to monitoring for drift and fairness over time, and steady emphasis on updated service names and builtain safety features across Azureas AI portfolio. The guidance continues to reward answers that combine effective capability with explicit risk controls. Now stitch the hour together so the essentials stick. You can recognize supervised and unsupervised learning, keep your data splits clean, and pick metrics that fit classification or regression. You can separate image classification, object detection, and document extraction and map them to Azure AI Vision, Custom Vision, or Azure Document Intelligence. You can turn audio into text, apply Azure AI Language for sentiment, key phrases, entities, PII detection, and summarization, and coordinate conve