Lost in Translation, Found in Tests

The situation

Code Nexus's creators don't live in one language. Some caption in isiZulu and English in the same sentence; some send support a voice note in French from a noisy venue. Code Nexus Assist now reads text, transcribes audio, translates and speaks back, and passes structured results — language, sentiment, entities — to the routing system. Last night the language vendor upgraded the model, and this morning the nightly tests are red. Nobody knows yet whether the meaning moved, the tone moved, or just the shape of the JSON.

What you'll learn

• AI-103 text and speech: language detection, sentiment, entities, structured JSON output contracts, translation and tone, transcription confidence, speech translation and synthesis as separate stations, audio reasoning and its limits, explicit error handling, regression tests with known-answer fixtures.
• Text-to-speech, custom speech models, speech as an agent modality, audio input to multimodal models and Azure Translator.

Who you work with

• 

  Jax Rivera

  Creator Partnerships Lead

• 

  Amara Okafor

  ML and Data Engineer

Scenes

1. 1.Tuesday night in Durban
2. 2.The contract and the examples
3. 3.Language flows that fail politely
4. 4.Speech translation, on its own station
5. 5.What did the upgrade change?
6. 6.Audio reasoning against known answers
7. 7.Language and speech check

What you leave with

You diagnosed the supplied contract regression and built simulated language, speech and evaluation flows. The lasting plan validates boundaries, checks meaning and tone, manages versions and reports failures explicitly. Confidence and exact-answer fixtures cover limited cases; they do not prove general language accuracy or audio-based mood inference.