TinkerChained: the ear writes a transcript, the model reads and writes text, the mouth speaks it.
Chapter 6 of 7
One model or three
A vendor pitches the team a speech-to-speech model: no ear, no mouth, one model that hears and speaks. The support lead asks what Larkspur would gain and what it would give up.
This chapter: When is a chained pipeline the better choice, and when is one speech-to-speech model?
You will learn
- What a speech-to-speech model does differently
- What the chained line gives you: a transcript to read and check between stages
- Why the turn-taking rules apply to both
Words you’ll meet (2)
- Speech-to-speech model
- One model that takes audio in and gives audio out, with no separate transcript step in the middle.
- Policy check
- A rule the host applies to text before acting on it, like the Guardrails module’s checks.
The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.
Step 1 of 3
A vendor offers Larkspur a speech-to-speech model: one model that takes audio in and gives audio out. The desk would lose its ear and its mouth. Before you switch: