In-car voice commands, before and after training

A small language model (Qwen3, 1.7 billion parameters) turns a spoken command into a vehicle function. It was tested on 76 commands it had never seen: once as downloaded, and once after training on the distil labs platform.

Before training 50% 38 of 76 commands right
After training 93% 71 of 76 commands right

Ten of the 76 commands

What was saidBefore trainingAfter training
“I’m cold, seat heating to three and ambient light to warm white.”from the front passenger seat Before trainingRightSeat heating level 3 for the front passenger, ambient light warm white. After trainingRightSeat heating level 3 for the front passenger, ambient light warm white.
“Heating on please.”from the front passenger seat Before trainingWrongSwitched on the seat heating without asking. After trainingRightAsked which heating was meant.
“Seat heating on high.”from the rear left seat Before trainingWrongSet level 5. The scale ends at 3. After trainingRightSet level 3 for the rear left seat.
“I’m freezing.”from the front passenger seat Before trainingWrongTurned the temperature down. After trainingRightTurned the temperature up for the front passenger.
“Close the windows.”from the driver’s seat Before trainingWrongClosed only the driver’s window, in a reply the car could not read. After trainingRightClosed all windows.
“Stop the navigation.” Before trainingWrongStarted a route to work. After trainingRightCancelled the navigation.
“Navigate to Schlossplatz 1 in Stuttgart.” Before trainingWrongStarted a route to work. After trainingRightSet the destination to Schlossplatz 1 in Stuttgart.
“Volume to forty.” Before trainingWrongTried to set volume 40. The scale ends at 30. After trainingRightAsked for a volume level, because 40 is off the scale.
“Open the glovebox.” Before trainingWrongOpened the trunk. After trainingRightDeclined. The car has no such function.
“hey um seat ventilation to one please”from the rear left seat Before trainingRightSet seat ventilation to level 1. After trainingWrongAsked for the level, although it was given.

Right means the function and every setting match exactly, or the model correctly asks back or declines.

Training took 1 hour 34 minutes and started from 96 hand-written examples.

The vehicle functions and all commands were made up for this demo. No customer data was used.

The before-training answers shown here come from a separate run of the same untrained model. Measured on 17 November 2026.