The kitchen and the recording
Try it
What was said, and what the room sounded like
Half past six on a Tuesday. The dishwasher is running a metre from the kitchen table, loud enough that everyone leans in a little. Nadia, who is seven, is telling you about Friday. Her grandmother is at the sink behind her with the tap on. A phone lies face up between the fruit bowl and a stack of mail, listening, writing down what it hears.
Here is what was actually said. The class is baking on Friday and every child brings one ingredient; Nadia’s is flour. There is a form that has to be signed first. She says her teacher’s name twice, clearly, both times: Mrs. Okafor. Partway through, the dishwasher reached its drain cycle and roared for about four seconds — you could not hear a word through it, and you asked her to say that part again. Her grandmother, who learned English as an adult, cuts in from the sink to ask about the Friday pickup.
That paragraph is standing in for the recording. There is no audio on this page, and it turns out not to matter, because everything worth judging here is in the text that came back. You were in the room. That is the only advantage you have over the transcript, and this demo is about spending it well.
What the sound turns into
Sound in
Two voices, one dishwasher
Speech model
One line of guessed text
The assistant
A spoken reply
Sound out
The sound stops at the second box. Everything after that point is working from a line of text, and that line is a guess — the assistant never hears the room, only the guess.