Take the noise out
DPDFNet is Ceva's speech cleaner for earbuds and calls. It hears 20 ms at a time and works out, frequency by frequency, what is voice and what is noise, carrying a memory from one moment to the next.
Pick a noise and how loud it is, then Clean it. Flip between Noisy and Clean while it plays, or tap the picture. Changing the noise cleans it again. The picture is a spectrogram: low sounds at the bottom, high at the top.
The voice is the Apollo 13 crew in 1970 (NASA, public domain), radio static and all. The added noises are made in your browser; the cafe chatter is that same clip, backwards and at other speeds. You can also use your own recording or file; it never leaves your device.
Runs on your device. Your voice is never uploaded. This is the 48 kHz model with 2 of its dual-path blocks.