profezzorn, I've been thinking and playing around a bit with error notifications.
Currently there are 3-4 errors that have talkie voice notifications. While it's cool to have voice messages with a pretty low memory footprint, they don't sound very good (and are often hard to make out) and the message itself is not exactly what the error is (ie Bank Open means a file was not found).
I wondered about putting recorded audio messages in program memory. Paul Stoffgren (Mr. Teensy) has done some work on this. Paul estimated that you could put 20 seconds of audio in program memory of a Teensy 3.2. I made 4 error messages using a test-to-speech website and came up with 4 seconds of total audio, so that would take 1/5 of program memory. Is this a direction you think is worth pursuing? The plus side is the audio error messages would sound better and be exactly what you want, but is it worth 1/5 of the program space? Let me know what you think.
Another option would be to put error audio files on the SD card. Of course this doesn't work for SD card or file errors. Perhaps beeps could be used if the SD card can't be read (ie 1 beep means low battery, 2 beeps means no SD card, 3 beeps means file not found, 4 beeps means bad blade). Or maybe just use beeps and no voice messages. Again, your thoughts?
Almost forgot to reply to this...
Let le's do some Math (tm).
We have 256kB of ROM. For error messages, phone quality would be sufficient, which is 8-bit 8kHz audio, or 8 kb / s, 256 / 8 = 32 seconds of audio.
I'd feel silly if I didn't apply some sort of compression to that though. A simple linear-predictive-encoding can double the amount of audio with relatively little loss in quality, which means 64 seconds of audio, or 4kb / s.
We probably only need 10 seconds of audio though, which would take up ~40kb. 40kb is not that much, but I can pretty much guarantee that it'll be the first thing to go if we start running low on space.
Another alternative would be to simply generate new error messages using the current speech synthesizer. The original talkie script comes with a matlab script that is supposed to be able to generate talkie data. Maybe it's possible to tweak that a little to make it generate higher quality. (Like, maybe by upping the sampling rate to 11kHz)