Between the document pipelines and the generative agents sits a family of services that quietly powers the most visible AI features users meet: speech recognition, synthesis, translation, and the language analysis toolkit. Today is a working tour organized around what you build with them, ending with the pattern that combines them all.

Speech: Both Directions

Speech to text handles real time streams and batch transcription, with diarization separating speakers, word level timestamps, and language identification for mixed language audio, the realistic condition in this region, where a sentence starts in Arabic and ends in English. Accuracy on your domain is a training decision: custom speech models fine tuned with your domain text and pronunciation lists lift recognition of product names, drug names, and jargon substantially, and the evaluation habit applies, a labeled test set of your actual audio, word error rate measured before and after, no anecdotes. Text to speech’s neural voices carry the other direction, with SSML controlling pace, emphasis, and pronunciation, styles for scenarios, and custom neural voice building a branded voice from studio recordings, gated behind an approval process for the obvious misuse reasons, use it for the IVR and product voice where brand justifies the investment, stock voices everywhere else. Speech translation chains recognition and translation in one stream for the live caption and multilingual meeting scenarios.

Language: The Text Toolkit

The language service bundles the analysis primitives: entity recognition and linking, key phrases, sentiment with opinion mining, summarization for documents and conversations, and custom classification and NER trained on your labels for the routing and extraction jobs that predate LLMs and still outprice them at volume, a classifier at fractions of a cent per document versus prompting a frontier model for the same routing decision is the FinOps posts applied to AI. PII detection deserves its own sentence: it identifies and redacts names, IDs, phone numbers, and financial details across dozens of languages, and it is the compliance workhorse, redaction before logging, before analytics, before content enters the RAG indexes from this week’s first post, the control that lets transcripts and tickets flow into data platforms without flowing personal data everywhere. Conversational language understanding covers intent and slot extraction where deterministic routing needs training data guarantees, and the honest modern guidance is a spectrum: fixed taxonomies and high volume favor the trained small models, open ended understanding favors the LLM with the guardrails from two days ago, and plenty of systems route with the former and respond with the latter.

The Call Center Pattern

The composition that exercises everything: audio from the Communication Services calling post streams to speech to text with diarization; PII redaction scrubs the transcript; the language service tags sentiment and entities while a generative model produces the summary and next actions through the Foundry agent patterns; results land in Event Hubs and flow to the operational dashboards and the lake for analytics; and the agent assist loop pushes suggestions back to the human agent through SignalR in real time. Every arrow in that sentence is a September or October post, which is the point of a series like this: the AI services are not a separate estate, they are workloads on the same private networking, managed identities, monitoring, and cost allocation as everything else, and they inherit the platform’s maturity the day they deploy. The invariants apply without exception, private endpoints, Entra auth, diagnostics on, containerized speech for the sovereignty cases that require processing in your boundary, and the usual monthly look at usage against commitment tiers because speech minutes at call center volume deserve the reservation math from the FinOps posts.

Cheers
Osama