I'd guess about 1 MB order of magnitude. That would be a butt-load even for samples.
1 MB would be enough for 45 seconds audio at 8 bit PCM @22 kHz. If they're half competent, they could use ADPCM or much better (and cheaper to decode too!) vector quantization.
With VQ, you could go even down to ~1 bits (~6 minutes of audio per megabyte) per sample while maintaining good audio quality. Decoding is very simple, so 6502 would have plenty of oomph to do that.
VQ is a very well known technique. For something quick, you might be interested to check this blog post about (unmodified) C64 playing high quality audio at 44.1/48 kHz, VQ compressed down to 96-128 kbit/s: https://brokenbytes.blogspot.com/2018/03/a-48khz-digital-mus...
the poster mentioned it was for greeting cards, I remember their used to be a lot of greeting cards that would play a recording when you opened them. I bet they loaded low-quality samples into the ROM for that.
The 6502 can only address 64K, so I'd assume 32K of ROM at 8000-FFFF with 128 (not 100) bytes of RAM mapped at both 80-FF (zero page) and 180-1FF (for the stack) and 128 MMIO bytes at 00-7F. It depends on how much they've modified the base 6502 design, though; the above is for "not at all".
In such an integrated design 6502 can address exactly as much as you want/need it to.
Just implement bank switching for top 32 kB of address space.
256 banks (= one 8-bit register) * 32 kB = 8 MB. Need more? Add another bank switch register and you got up to 2 GB.
1 MB would be enough for 45 seconds audio at 8 bit PCM @22 kHz. If they're half competent, they could use ADPCM or much better (and cheaper to decode too!) vector quantization.
With VQ, you could go even down to ~1 bits (~6 minutes of audio per megabyte) per sample while maintaining good audio quality. Decoding is very simple, so 6502 would have plenty of oomph to do that.