My guess is you've edited the audio lib PWM object, for a divisor of only 50 from the bus clock (48 MHz / 50 = 960 kHz). Teensy can easily produce PWM carriers in the AM radio band. It's pretty amazing the PCB trace with PWM can radiate enough RF for the radio to pick up, since it's on top of a ground plane. Even a short wire extending from the pin would probably help quite a lot.
Of course, the data scaling code must be changed too, so the 16 bit data maps into the limited 0-49 range... perhaps only a small part of the 50 available codes, so the PWM waveform stays closer to 50%. I would imagine if you get too close to either duty cycle extreme, the amount of radiated energy in the carrier becomes quite a bit less.
Maybe you just mapped the signal onto the raw codes with each sample repeated many times. Or maybe you're using another timer to trigger the DMA, since there's no need to rewrite the CV register repeatedly with the same value. Or maybe (and I hope) you might have kept the update at 1 sample per PWM waveform done some really crafty dithering and/or noise-shaped upconversion from 44.1 kHz to 960 kHz? My guess is such conversion might be able to achieve pretty good quality, despite the quantization from 16 to approx 5.6 bits (or less, if not using the entire 0-49 range), especially with sigma-delta noise shaping. From the video, it's hard to tell if the poor quality is due to weak RF, only about 5 bits of sample res, or the radio's speaker, or recording equipment used.