To handle decoding and the overhead of moving bits around, it uses a dual-buffer DMA approach. It reads data from an SD card, running the decoding routine on the main CPU. The decoded samples are written into one of two buffers. While one buffer is being written to, the other is read via DMA and sent to the PWM buffer (more on PWM audio here: https://x.com/beala/status/1826147833168408738). Because DMA is used, data transfer can occur in parallel with decoding, without involving the main processor. When the bits have all been written to the PWM buffer, an interrupt triggers a buffer swap, and the process repeats.
I'm leaving out many details. Tuning the pwm speed to achieve correct playback speed was quite an adventure. And when I needed to increase the clock speed to play higher sample rates, this caused I/O errors with the SD card that I never managed to solve. Although I learned a lot from this process, I ultimately opted to use an MP3 module in my final build: https://store-usa.arduino.cc/products/dfplayer-a-mini-mp3-pl...
This is probably routine for an experienced embedded engineer, but it was quite a learning experience for me!