audioeffects.deesser

The dbx 902 de-esser: sibilance measured against the programme, not against a threshold knob.

The player’s text is the class docstring; the CLAIMS table in the class’s test file ties its claim sentences to tests.

class audioeffects.deesser.DeEsser(source, *options, **keywords)[source]

Bases: Component

De-essing by comparing the sibilant band with the programme.

No threshold. It listens to how loud the “s” band is against the whole track and ducks only when the band comes too close, so one setting works on a whisper and a belt. At every shipped patch’s Sensitivity the reduction agrees within 1 dB from -6 dBFS down to -55 dBFS, and at any Sensitivity within 0.5 dB from -6 dBFS to -40 dBFS. On the quietest takes, around -55 dBFS, other Sensitivity settings can drift by more than 1 dB.

Frequency is where the “s” lives, and it moves the crossover and the detector together. Range is the most it may take away, and it never takes more than 0.05 dB past it. Range 0 is a wire, byte for byte. The reduction approaches Range without reaching it: at Range 12 the deepest any Sensitivity gets is more than 0.5 dB short. Sensitivity is how close to the full-band level the band may come before it counts.

Mode. Broadband ducks the whole signal, for a vocal: the high and low bands move within 0.5 dB of each other. HF Only ducks the high band alone, for pick noise and cymbal edge, and the low band stays within 0.25 dB. With nothing to duck, HF Only’s split sums back flat within 0.25 dB.

Release is how fast the reduction lets go. The default 3.5 ms recovers at the dbx 902’s own 925 dB a second, within 10 %, and the rest of the knob moves the rate well away from it. Attack gets shorter as the “s” gets louder. Listen plays the band the detector hears. The Effect controls page lists what the knobs span and the patches.

Latency is zero, and no setting adds any. The tail reaches exact zero within 256 samples. It needs audiobiquad, audiodynamics and audioroute, so a stock CircuitPython board can’t run it.

LAZY_BUILD = True

Construction arms the voices on silence (Component._play_voice), so a Mix 0 set before the first pull hands out the source whole.

NAME = 'DeEsser'

The stable provider name audioeffects.create() discovers.

TIER = 'audiodsp'

STOCK or AUDIODSP, with REQUIRES naming the audiodsp-own modules an AUDIODSP class needs.

Type:

roadmap §3

CAPABILITIES = ()

sibilance has nothing to do with tempo. Nothing in this class reads self._transport().

Type:

(), and the dossier’s reason

LATENCY_SAMPLES = 0

A VCA under a detector. The IIR split delays phase, never onset, and no look-ahead is offered.

TAIL_SAMPLES = 256

One node block, as a bound on what was measured. Broadband mode has no tail at all - the audio never enters a filter - and HF-only mode’s longest is the Butterworth pair at the bottom of the Frequency span: 170 samples at a 800 Hz corner at 48 kHz, 51 at 2.5 kHz, 14 at 8 kHz, every one of them reaching exact zero, which is what audiobiquad exists for. The numbers are in the evidence pack.

After a reset, a rejoin or a control move the detector can leave the output under 16 LSB from an untouched copy for up to three blocks, which tail_samples does not bound. See “Detector and coupling state” in docs/reference/component-api.md.

WALKS_TO_BYPASS = True

Range 0 walks the output mixer to the dry before the port switches.

MACRO_LABELS = ('Frequency', 'Range', 'Sensitivity', 'Mode', 'Release', 'Attack', 'Listen')

Knob names, one per macro index; one mode per label; the private engineering spans behind them; and the patches on the 0-127 grid.

reset()[source]

Come back as a fresh instance at patch 0: the graph is built again and the old one released.

A reset that cleared this graph in place never came back fresh. A Splitter’s ring cannot be cleared - the type exposes tap() and nothing else, and SplitterTap.reset_buffer is “deliberately nothing” - so the old reset pushed eight taps empty with audiocore.get_buffer (compiled out of CircuitPython’s board builds, where it raised), and with the rings empty it still differed from a fresh instance from the first frame of material on. A graph built again is a fresh instance, by construction, on every interpreter. The output port is re-pointed at the new tail before the old graph is released, so a pump holding the port never reads a released node. The borrowed source is not reset or rewound; it plays on from where it stands, as a fresh instance handed it would.

gain_reduction_db()[source]

What the gain cell did to the last frame, in dB.

This is the cell’s own number, before the Range blend, so it is the detector’s reading rather than what reached the output. The output’s reduction is 20*log10((1 - alpha) + alpha * 10**(cell/20)).