Sampled sound

References

1. ^The Game Boy has four sound channels, each specialized, with two pulse channels with four fixed duty settings, a wave channel that plays a user-supplied wave, and a noise channel.WikiPan Docs, “Pan Docs: Audio Overview”
   How we know

   Channels 1 and 2, the “pulse channels”, produce pulse width modulated waves with 4 fixed pulse width settings. Channel 3, the “wave” channel, produces arbitrary user-supplied waves. Channel 4 is the “noise” channel, producing a pseudo-random wave.

   Checked against the evidence by its author, an AI model, 2026-10-21.

   Checked by a second AI model, 2026-10-21.

   In the claim ledger

2. ^ abThe wave channel loops 32 4-bit samples from wave RAM ($FF30-$FF3F, 16 bytes, upper nibble first), clocked at 2,097,152 Hz divided by $2048-p$, so a full wave repeats at $65536/(2048-p)$ Hz. Its output level is 100%, 50%, 25%, or off.WikiPan Docs, “Pan Docs: Audio Registers”
   How we know

   Wave RAM is 16 bytes long; each byte holds two “samples”, each 4 bits. As CH3 plays, it reads wave RAM left to right, upper nibble first.

   The wave channel’s period divider is clocked at 2097152 Hz, once per two dots, and its waveform is 32 samples long.

   100% volume (use samples read from Wave RAM as-is)

   It’s also possible to artificially “increase” the wave’s length by loading a new wave as soon as the whole buffer has been read; this is sometimes used for full-on sample playback.

   Checked against the evidence by its author, an AI model, 2026-10-21.

   Checked by a second AI model, 2026-10-21.

   In the claim ledger

3. ^The delta modulation channel (DMC) plays 1-bit delta-encoded samples from $\$C000+64A$, $16L+1$ bytes long, at one of 16 rates; a program can also load its 7-bit output level directly through $4011.WikiNESdev Wiki, “APU DMC”
   How we know

   The NES APU's delta modulation channel (DMC) can output 1-bit delta-encoded samples or can have its 7-bit counter directly loaded, allowing flexible manual sample playback.

   Sample address = %11AAAAAA.AA000000 = $C000 + (A * 64) … Sample length = %LLLL.LLLL0001 = (L * 16) + 1 bytes

   For example, on NTSC (1.789773 MHz), a rate of 428 gives a frequency of 1789773/428 Hz = 4181.71 Hz.

   Checked against the evidence by its author, an AI model, 2026-10-21.

   Checked by a second AI model, 2026-10-21.

   In the claim ledger

4. ^Sound runs on its own processor, the SPC700 (S-SMP) at 1.024 MHz with 64 KB of audio RAM, beside the S-DSP, which plays 8 voices from BRR samples at variable rates and outputs 16-bit stereo at a nominal 32,000 Hz.WikiSNESdev Wiki, “S-SMP”
   How we know

   Sony SPC-700 processor, clocked at 1.024MHz … 8 voice channels Decodes BRR samples at a variable sample rate using 4-point Gaussian interpolation … Audio-RAM: 64KiB RAM from two 32K x 8bit PSRAM chips. … Digital to analog converter, outputting the analog audio signal from the DSP's stereo 16-bit 32000 Hz digital audio.

   Checked against the evidence by its author, an AI model, 2026-10-21.

   Checked by a second AI model, 2026-10-21.

   In the claim ledger

5. ^BRR stores samples in 9-byte blocks of 16 4-bit samples, each block led by a byte with a shift, a decoding filter, and loop and end flags.WikiSNESdev Wiki, “BRR samples”
   How we know

   BRR is composed of 16-sample blocks, each of which is stored in 9 bytes of data. This is a 1 byte control block, followed by 8 bytes containing 16 4-bit samples to be decoded (high nibble first). … Left-shift (S), decoding filter (F), loop (L), end (E).

   Checked against the evidence by its author, an AI model, 2026-10-21.

   Checked by a second AI model, 2026-10-21.

   In the claim ledger

6. ^ abChannel 3 plays one of six 32-step waveforms, numbered 0 to 5. A song's note_type command picks one with its low nibble, and on each note the engine stops the wave channel, copies the 16 bytes into wave RAM, and starts it again.Codepret/pokered
   How we know

   3	dw .wave0
   ⋮
   8	dw .wave5 ; used in the Lavender Town and Pokemon Tower themes
   ⋮
   13; these are the definitions for the channel 3 instruments
   14; each instrument definition is made up of 32 points (nibbles) that form
   15; the graph of the wave
   16; the current instrument is copied to AUD3WAVERAM
   

   344	cp CHAN3
   345	jr z, .musicChannel3
   ⋮
   350.musicChannel3
   351	ld hl, wMusicWaveInstrument
   352.channel3
   353	ld a, d
   354	and $f
   355	ld [hl], a ; store low nibble of param as wave instrument
   

   906Audio1_ApplyWavePatternAndFrequency:
   ⋮
   920	ld a, [de]
   ⋮
   924	ld hl, Audio1_WavePointers
   ⋮
   929	ld hl, AUD3WAVERAM
   930	ld b, AUD3WAVE_SIZE - 1
   931	ld a, $0 ; stop hardware channel 3
   932	ldh [rAUD3ENA], a
   933.loop
   934	ld a, [de]
   935	inc de
   936	ld [hli], a
   ⋮
   941	ld a, AUD3ENA_ON ; start hardware channel 3
   ⋮
   

   Checked against the evidence by its author, an AI model, 2026-10-21.

   Checked by a second AI model, 2026-10-21.

   In the claim ledger

7. ^ abPikachu's voice clips are 1-bit recordings played through the wave channel. The routine disables interrupts, fills wave RAM with $FF so the channel outputs a constant high level, and then for each bit sets channel 3's output level to 100% ($20) or mute ($00), with counted delay loops setting the rate.Codepret/pokeyellow
   How we know

   32	ld hl, _AUD3WAVERAM
   33	ld de, wSavedAudioWavePattern
   34.saveWaveDataLoop
   35	ld a, [hl]
   36	ld [de], a
   37	inc de
   38	ld a, $ff
   39	ld [hli], a
   ⋮
   54	ld a, $ff
   ⋮
   56	ld a, $20
   57	ldh [rAUD3LEVEL], a
   

   11.loop
   12	ld a, [hli]
   13	ld d, a
   14	ld a, $3
   ⋮
   16	dec a
   ⋮
   19REPT 7
   20	call LoadNextSoundClipSample
   21	call PlaySoundClipSample
   22ENDR
   ⋮
   24	call LoadNextSoundClipSample
   ⋮
   28	jr nz, .loop
   ⋮
   31	ret
   ⋮
   

   Checked against the evidence by its author, an AI model, 2026-10-21.

   Checked by a second AI model, 2026-10-21.

   In the claim ledger

8. ^ abThe sound engine runs on the SPC700. Its set-instrument command ($DA) reads a 5-byte entry from a table in audio RAM at $5F46 and writes the first four bytes to the voice's DSP registers, the sample number (SRCN), ADSR1, ADSR2, and GAIN, and keeps the fifth as a pitch multiplier, so every instrument is a recorded sample with an envelope.CodeSMWDisX
   How we know

   3SPC700Engine:
   ⋮
   7    arch spc700
   

   1174VCmd_DA:
   1175    CALL NextVCmd46                           ; vcmd DA: set instrument;
   ⋮
   1182    MOV Y,#$05
   1183    MOV.B ARam_14,#$46
   1184    MOV.B ARam_15,#$5F
   ⋮
   1186APU_0D56:
   1187    MUL YA                                    ; set sample A in bank at $14/15 width Y;
   ⋮
   1197    OR A,#$04                                 ; voice X SRCN;
   ⋮
   1211    CMP Y,#$04
   1212    BNE -                                     ; set SRCN, ADSR1/2, GAIN;
   1213    MOV.B A,(ARam_14)+Y
   1214    POP X
   ⋮
   

   Checked against the evidence by its author, an AI model, 2026-10-21.

   Checked by a second AI model, 2026-10-21.

   In the claim ledger