Monorepo for Aesthetic.Computer
aesthetic.computer
native: reduce dither to ±1 LSB — ±160 was audible as 24kHz fizz
At S32_LE, the ±160 int16 dither became ±10M int32 (~-46 dBFS)
which the MAX98360A amp reproduced as an audible high-frequency buzz.
Drop to ±1 LSB (±65536 int32 = -90 dBFS) — still prevents silence-
detect amp shutdown but completely inaudible.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
+8
-13
+8
-13
fedac/native/src/audio.c
+8
-13
fedac/native/src/audio.c
···
1494
1494
buffer[i * 2] = (int16_t)(mix_l * 26000);
1495
1495
buffer[i * 2 + 1] = (int16_t)(mix_r * 26000);
1496
1496
1497
-
/* DAPM keepalive: inject low-level dither when the buffer
1498
-
* would otherwise be silent. SOF DSP's internal silence
1499
-
* detector uses an RMS threshold (not just zero-detect)
1500
-
* and powers down the SSP1 BE DAI, dropping MAX98360A
1501
-
* sdmode to 0. ±1 LSB extended amp-on from 23s to 96s
1502
-
* but wasn't enough. ±160 LSB is -44 dBFS — still well
1503
-
* below audibility at normal speaker levels (~0.6% of
1504
-
* full scale) but gives the PGA enough RMS energy to
1505
-
* keep the pipeline active. The alternating pattern
1506
-
* averages to DC=0 so there's no pop/click risk. */
1497
+
/* DAPM keepalive: inject ±1 LSB when the buffer would
1498
+
* otherwise be all zeros. At S32_LE (after <<16) this
1499
+
* becomes ±65536 ≈ -90 dBFS — utterly inaudible.
1500
+
* Previous ±160 was audible as 24kHz fizz through the
1501
+
* amp. ±1 is enough to prevent the SOF DSP silence
1502
+
* detector from powering down the SSP1 BE DAI. */
1507
1503
if (buffer[i * 2] == 0 && buffer[i * 2 + 1] == 0) {
1508
-
int16_t d = (i & 1) ? 160 : -160;
1509
-
buffer[i * 2] = d;
1510
-
buffer[i * 2 + 1] = -d;
1504
+
buffer[i * 2] = (i & 1) ? 1 : -1;
1505
+
buffer[i * 2 + 1] = (i & 1) ? -1 : 1;
1511
1506
}
1512
1507
1513
1508
// HDMI audio: 1-pole low-pass filter + downsample