Finally, the PSG (Programmable Sound Generator AY-3-8910) is coming to life. It was harder than I thought. But it now has the 3 channels for tone or noise, with independent frequency and volume control. Only the the envelope generator is not present yet [edit: a month later, envelope was also working]. The sound is generated by microcode, no actual AY-3-8910 is used, but the sound is controlled in the same way as on a real sound chip.

It generates the typical sound of the eighties computer games.

It can now provide sound for the SymbOs operating system !  And here you see the first Youtube video of Isetta running SymbOs (made by Prodatron):

(The start-up screen of Isetta will have to be modified, it's quite a mess at the moment.)

A new version of the Isetta programming manual is now in the file section. It includes the commands to control the sound generator.

OPERATION

There are four 1 kB sound buffers, one for each of the three channels, and a fourth one for noise.

The squarewave sounds are generated in an unrolled loop. So if a wave has 5 samples 'high' and 5 samples 'low', to generate one period there will be two sequences that each put samples in the buffer, with each having five "(pc++) <- T" micro-instructions in a row. (Note the program counter is used as auto-incrementing pointer to the buffer).

The value in T is the required amplitude for this sound (positive value for high samples, negative value for low samples). So there is no separate volume-control stage, like the real PSG has.

At each scanline interrupt, a byte-sized sample is taken from each of the four buffers, the four samples are added and sent to the D/A converter (resistor array RN1). So the sample frequency is the same as the VGA line frequency, 31.25 kHz.

To control this, there are two Z80-output ports. One to select a AY-3-8910 register, and another one to write a byte to that register. The Z80 output ports can also be used by the 6502 processor.

EFFICIENCY

For each channel, producing samples takes around 3 cycles per sample. (The unrolled loop produces 8 samples on average, single cycle per sample, with also 16 cycles overhead).
For a full 525-line frame, this is 1575 cycles per channel. Noise has it's own channel, so there are 4 channels, together taking 6300 cycles each frame.

In each frame, 120 lines are for the CPU, minus interrupt it's around 350 usable cycles per line. Per frame, this is 120 * 350 = 42000 available cycles. So the sound takes 15% of the CPU time. For frequencies higher than 1kHz, there are just a few samples produced in each loop, so there is more overhead.

LINKS

There are several very nice articles about sound generation.

Want to know everything about synthesizers:

https://www.soundonsound.com/series/synth-secrets-sound-sound

Building the AY-3-8910 fully from TTL chips:

https://github.com/mengstr/Discrete-AY-3-8910

Very good description of the AY-3-8910:

https://www.cpcwiki.eu/index.php/PSG

Article about Linear-feedback shift registers, used to generate noise:

https://en.wikipedia.org/wiki/Linear-feedback_shift_register

The Isetta noise generator comes from this article. It's a 16-bit Galois type, quite simple:

       unsigned msb = (int16_t) lfsr < 0;   /* Get MSB (i.e., the output bit). */
        lfsr <<= 1;                          /* Shift register */
        if (msb)                             /* If the output bit is 1, */
            lfsr ^= 0x002Du;                 /*  apply toggle mask. */

Sound generation in the Novasaur TTL computer:

https://hackaday.io/project/164212-novasaur-cpm-ttl-retrocomputer/log/184708-roll-your-own-sid-chip

Interactive demo, shows harmonics of square waves and triangle waves:

https://www.desmos.com/calculator/emfmoefwqu

A suitable IC for driving a small speaker: LM4862, info HERE. Or use LM4991 (cheaper). Small Speaker could be 665-ASE03008MRLW150R. Or 179-CMS-30204-18L250. Or 497-SM500308-1.