Speech-to-Speech Model or Cascaded Pipeline? Decide for a Bank and a Language Tutor

Two teams ask for your recommendation.

• Bank. A voice assistant on the phone that checks balances, disputes transactions and issues card replacements. Regulators require an auditable record of what was said and approved, tool calls must follow strict policy, and PII must be redacted from stored data.
• Language tutor. A consumer app where learners converse freely, the tutor must react to pronunciation, hesitation and tone, interruptions are frequent, and there are no irreversible actions.

1. State the architectural difference and what each approach gives up.
2. Recommend for each team with reasons; consider a hybrid.
3. What would you measure in a pilot to confirm the decision?