Why the first second decides a voice lesson
A spoken lesson is judged by the pause before the tutor answers and by whether it can be interrupted. We treat both as hard constraints, not tuning.
When a student finishes a sentence and nothing happens for three seconds, the lesson is over in their head, whatever the tutor says next. That pause is the product. Everything else — the curriculum, the avatar, the report — is judged through it.
A budget, not a benchmark
We give one turn a budget of two seconds and aim for one and a half: from the moment the student stops talking to the first word back. Not the whole answer — the first word. Most of the budget is spent in four places, the network and the avatar take the rest, and each stage streams into the next:
- Endpointing: deciding that the student has stopped, not paused to think.
- Speech recognition finalising the transcript.
- The tutor model producing its first words.
- Speech synthesis producing the first audio chunk.
None of them is allowed to wait for the previous one to finish. The model starts on a partial transcript when the endpoint is confident; synthesis starts on the first sentence, not the whole answer. Latency marks on every stage go into the logs, so a regression shows up as a number, not as a feeling.
Interruption is the other half
A tutor that cannot be interrupted is a recording. When the student speaks over it, three things have to happen at once: synthesis stops, the audio already buffered is dropped, and the generation that was still running is cancelled. Miss one and the tutor "finishes its thought" while the student is already talking — which is what every chatbot with a voice does.
We call it barge-in, and it is a requirement of the engine, not an option a product can turn off.
What this means for a partner
If you order a tutor for your subject, you are not ordering prompts. You are ordering a pipeline that keeps those two promises under load, in your language, with your voice provider. The curriculum is data; the promises are the product.