Understand the responsibilities behind a spoken conversation.
1. STT, model and TTS
Speech-to-text turns audio into text, the model interprets the request, and text-to-speech produces audio. Each stage can fail independently and adds latency.
2. Pipeline versus realtime
A staged pipeline offers explicit intermediate text and provider boundaries. A realtime speech model can integrate stages but still requires session, interruption and permission handling.
3. Task boundaries
Choose a narrow voice task with a text alternative. Speech is an input and output channel, not authority to perform unconfirmed sensitive actions.
Worked scenario
A note assistant transcribes a shopping item and asks for confirmation before saving. A transcription mistake should not immediately change stored data.
Apply it
Draw both a staged and realtime architecture for the same task and compare failure visibility.
Check your understanding
You can name each latency and correctness boundary. Explain the decision and show evidence from your implementation or design. If you cannot demonstrate it yet, revisit the relevant section before continuing.