I already spotted it but the quality is not what I expect (not trying to diminish Pacman's work here!) and since I can read the text and heard it, "priming" must intervene here. To judge the real quality, you need to hear it without knowing what it says - which means having someone else with closed eyes and just hearing a sentence you typed, and asking for this person to tell you what he/she heard.
Yes indeed. I'm sure the 'priming' is a big factor. I remember a VIC-20 demo where speech was created in the music but you wouldn't really be able to understand it unless you read the words which were also on the screen.
I fully agree with this: I had noticed that the PWM generated alone without the text is difficult to understand. However, I applied bandpass filters on the original soundtrack, limiting the maximum frequency to 2kHz, but maybe I would have to decrease the bandwidth further. I'm in the middle of trying to improve the quality, but in PWM it's really not easy!