Text and speech remain core AI workloads in the current AI-901 blueprint, but the April 2026 objectives frame them as capabilities that candidates should both recognize and implement in lightweight Microsoft Foundry solutions. The text objectives include keyword extraction, entity detection, sentiment analysis, and summarization. The speech objectives include recognition, synthesis, spoken prompts with multimodal models, and Azure Speech in Foundry Tools.<\/p>\n
For candidates preparing for the AI-901 exam<\/a>, the easiest way to organize this material is by the direction of information flow. Text analysis starts with language and extracts meaning. Speech recognition converts spoken audio into text or another machine-usable representation. Speech synthesis converts text into spoken audio. A multimodal interaction can combine speech, text, images, and conversation in one experience.<\/p>\n Organizations have large amounts of text in support tickets, reviews, messages, documents, and forms. Text analysis helps turn that unstructured language into structured information an application can use. The current AI-901 study guide explicitly names keyword extraction, entity detection, sentiment analysis, and summarization.<\/p>\n Keyword extraction identifies important terms or phrases. Entity detection recognizes named things such as people, organizations, locations, products, or other domain-relevant entities. Sentiment analysis estimates the expressed attitude or polarity of text. Summarization reduces longer content into a shorter representation that preserves the important ideas.<\/p>\n These techniques solve different problems. A support dashboard may use sentiment to prioritize unhappy customers, entity detection to identify products mentioned in tickets, keywords to group recurring issues, and summarization to help an agent review a long conversation quickly.<\/p>\n AI-901 scenarios often contain extra detail, but the required output usually reveals the workload. If the application must identify the company names mentioned in a document, entity detection is the obvious direction. If it must determine whether feedback is positive or negative, sentiment analysis fits. If it needs the main ideas from a long document, summarization is more appropriate.<\/p>\n Do not choose a technique because a keyword in the question sounds familiar. Ask what transformation must happen between input and output. The same input text can support several techniques, but only one may satisfy the business requirement.<\/p>\n This reasoning also prevents overengineering. A small extraction task may not need an agent or elaborate generative workflow. The simplest capability that satisfies the requirement is often the best fundamentals answer.<\/p>\n Speech recognition is commonly described as speech to text. The system receives spoken audio and converts it into a text representation that can be searched, analyzed, stored, translated, or sent to another AI capability. Common use cases include transcription, voice commands, call analytics, and spoken interaction with applications.<\/p>\n Recognition quality depends on real-world conditions such as noise, microphone quality, language, accent, specialized vocabulary, and overlapping speakers. A production system should therefore be evaluated on representative audio rather than only clean studio samples.<\/p>\n At AI-901 level, the key is to identify the direction correctly. If the user speaks and the application needs text, recognition is involved. If the application already has text and must speak it to the user, that is synthesis instead.<\/p>\n Speech synthesis, or text to speech, generates audio from text. It is useful for accessibility, virtual assistants, navigation, automated notifications, reading applications, and any workflow where spoken output is more useful than a screen.<\/p>\n A good synthesis experience is not simply \u201caudio exists.\u201d Voice quality, language, pronunciation, pace, and the context in which speech is delivered all affect usability. In an accessibility scenario, the application should also preserve alternatives rather than assuming audio is appropriate for every user.<\/p>\n The recognition-versus-synthesis distinction is one of the simplest exam patterns, but it is foundational because many real applications use both directions in the same conversation.<\/p>\n The current study guide explicitly includes responding to spoken prompts using a deployed multimodal model. That reflects the wider shift from isolated speech pipelines toward models that can handle several modalities as part of one interaction.<\/p>\n In a conversational application, the user may speak a request, the system may interpret the audio directly or through speech processing, and the model may combine that request with text, images, or other context. The output could be text, audio, or both depending on the application.<\/p>\n The broader multimodal AI fundamentals<\/a> are useful here because speech is no longer an isolated feature. The exam expects candidates to recognize when multiple forms of input belong in the same solution.<\/p>\n The implementation objective also includes building a lightweight application with Azure Speech in Foundry Tools. At fundamentals level, think in components: the application captures or supplies audio, calls the speech capability, receives the recognized or synthesized result, and then uses that result in the user experience.<\/p>\n A voice-enabled support tool could recognize a caller\u2019s speech, pass the text into a downstream workflow, and synthesize a response. A transcription utility could stop after recognition and save the text. A reading assistant could start from text and use synthesis only.<\/p>\n The implementation should match the scenario. Avoid adding a large generative layer when the requirement is simple transcription, and avoid forcing a narrow speech API to solve a richer multimodal reasoning problem when a deployed multimodal model is the better fit.<\/p>\n Consider a call-center application. Speech recognition converts the conversation into text. Text analysis identifies sentiment, products, and key issues. Summarization creates a concise case note. The application can then present the result to a human agent or generate a spoken response if the workflow allows it.<\/p>\n Each component has a clear role, and the pipeline can be evaluated at each boundary. If the transcript is wrong, downstream sentiment and summarization will also suffer. If recognition is accurate but the summary omits the customer\u2019s main problem, the failure is in the text-analysis or generative stage rather than the speech stage.<\/p>\n Voice applications also make the direction of data movement important. Recognition is an input capability: audio becomes text so that downstream language processing can work with it. Synthesis is an output capability: text becomes audio so that a user can hear a response. A conversational experience may use both, but they solve different problems and should be tested separately. Recognition quality can be affected by noisy audio, unclear speech, or the language being used, while synthesis quality is judged by whether the generated voice communicates the intended text naturally and intelligibly. At the fundamentals level, candidates should be able to look at a requirement such as transcription, spoken command handling, accessibility playback, or voice response and identify which direction of transformation the solution actually needs.<\/p>\n This dependency thinking is useful for AI-901 because it helps candidates reason about multi-capability scenarios without treating the entire system as one opaque \u201cAI service.\u201d<\/p>\n Speech recordings can contain personally identifiable information, biometrics, health details, financial data, or private conversations. Text can reveal the same information after transcription. Privacy and security controls should therefore follow the data through the complete workflow rather than protecting only the original audio file.<\/p>\n Fairness can appear when recognition quality differs across accents, dialects, or demographic groups. Inclusiveness matters when the experience supports users with different abilities or communication needs. Transparency matters when users should know that calls are being transcribed or analyzed. Accountability matters when automated analysis influences a consequential decision.<\/p>\nText analysis converts unstructured language into useful signals<\/h2>\n
Choose the technique from the desired output<\/h2>\n
Speech recognition turns audio into language the application can process<\/h2>\n
Speech synthesis turns text into spoken output<\/h2>\n
Spoken prompts can be handled through multimodal models<\/h2>\n
Azure Speech in Foundry Tools supports lightweight speech applications<\/h2>\n
Text and speech can be combined into one practical workflow<\/h2>\n
Responsible AI applies to language and audio data<\/h2>\n