In VirtuaNES there is a file tells the user that he can add custom voice samples to Moero!! Pro Yakuu. Without them the game stays silent. In Bases Loaded, the NES version of the game, there are sound samples for the umpire's calls. I remember hearing something about some kind of extra Famicom sound that remained mysterious to everybody on these boards. Could this be it? Emulating speech synthesizers is quite different from emulating sound chips.
I just opened Jaleco's Bases Loaded (U) in a wave editor, and it appears to use DMC samples at rate $F (roughly 33.5 kHz). SCAT and Big Bird's Hide and Speak both use raw samples, stored compressed in the ROM and decompressed by the NES into timed writes to $4011. Both the NES and the Famicom can play DMC and raw samples.
Famicom "expansion" sound is synthesized by circuitry inside the mapper, as the Famicom has an extra pair of pins for audio, which the NES lacks. This synthesis may come from sample data in a second ROM unconnected to the PRG bus. (EDIT: Yes, that's true in this case.)
Kevin Horton has that particular Famicom cartridge. and it has a single small chip in it which handles the extra sound. Clocked on a ~600kHz crystal, it simply contains sixteen different PCM samples. Perhaps, one of these days, he'll extract the digital data out of it, most likely by clocking it manually and reading the values through an ADC.
Cartridge
PCB
So essentially there is a DAC on the cart that transforms the data into analog waveforms and sends it along with the Famicom's internal sound to the audio outputs. Or is the chip a simple device that when written to sends its data to the PCM channel in the Famicom itself? After all, it is sort of like a DAC.
The sound ROM and DAC is all inside a single chip, using 6 data lines to control it (4 to select sound, another to start, and another to loop)