A software clone of the Inovonics OmegaFM FM broadcast audio processor, written in Python. It is built like real broadcast gear — a complete filtering and dynamics chain, an all-digital stereo/MPX generator, and a standards-compliant RDS encoder — behind a classic hardware front panel that scales to any screen.
Audio is processed at 48 kHz in the multiband section, then 4x-oversampled to 192 kHz for everything that shapes peaks:
IN -> input gain -> phase rotator (Optimod-style, 4x AP @200 Hz)
-> broadcast-grade gated slow AGC (fast-detector gate that
freezes gain AND its loudness integrator rock-solid across gaps,
single-speed slew - rides identically at any input drive, never
pumps) (±12 dB gain riding)
-> 4-band parametric EQ (100 / 400 / 2k / 8k)
-> psycho-acoustic bass (sub harmonics + low shelf; factory presets
keep total LF EQ boost <= 1 dB - perceived bass comes from the
harmonics, not from level)
-> Linkwitz-Riley 4 crossover @ 160 / 1k / 5k (allpass-compensated,
band sum verified flat to < 0.01 dB)
-> 4-band spectrum-consistency leveler: true 300 ms integrated
loudness detection, ±2 dB correction window, and *mix-protected*
correction - the bands are zero-mean (net gain always 0 dB; the
AGC owns level) and coupled to within ±2 dB of the pack, so it
corrects spectral shape gently and can never repaint a mix
(validated on strings-dominant classical: no band fades, no bass
creep, ≤4 dB total shape authority)
-> 4-band soft-knee compressor (the "signature") with
program-dependent release: GR releases only down to a slowly
gliding platform, never collapsing to unity between hits
-> band mix -> drive
== 4x polyphase upsample to 192 kHz ==
-> true bilinear 75/50 µs pre-emphasis (accurate to 0.02 dB @ 15 kHz)
-> dynamic HF limiter (protects the emphasized top end)
-> wideband look-ahead limiter (2 ms LA, gain-riding release:
a platform that charges down in ~0.25 s and bleeds up at only
0.8 dB/s, with a ~90 dB/s fast return *to the platform*)
-> soft-knee clipper (loudness)
-> 15 kHz linear-phase brick-wall (Kaiser, >80 dB) - pilot protection
-> safety clip
Stereo generator — one phase accumulator drives the 19 kHz pilot,
the 38 kHz DSB-SC subcarrier (sin 2θ) and the 57 kHz RDS carrier
(sin 3θ), so all three are phase-locked by construction.
The filtered composite clipper clips only the audio MPX, band-limits
the clip products with a linear-phase 53 kHz LP + 19 kHz notch mask,
then injects pilot and RDS after the mask using a delay-compensated
phase — measured stereo separation: 75 dB through the full composite
path (hardware spec: > 65 dB). A "raw" mode (clip everything, then LP)
and "off" are on the panel too.
RDS encoder — real EN 50067 groups: 0A (PS, DI stereo, no-AF) and
2A (64-char RadioText with A/B flip), CRC-10 checkwords with offset
words, differential + biphase coding, raised-cosine (β = 1) pulse
shaping placed at exact fractional-sample symbol instants. PS refreshes
~0.7 s, RT ~2.8 s. A watched now-playing text file can update RT/PS
live (RT=... / PS=... lines, or plain text → RT), with optional
Dynamic PS that rotates the RadioText through the 8-character PS field.
The whole gain chain rides - including the compressor. Every release in the dynamics chain returns to a platform, not to unity: the multiband compressor holds 87 % of its working GR through the gaps in percussive material and the WB limiter holds 89 % (both validated), so the final limiter glides with the compressor instead of grabbing on every syllable - while the attack shaping (the density character) is untouched.
The AGC and per-band leveler measure true integrated loudness (power-domain 300–400 ms ballistics) and only correct outside a deadband window around target, so their gains hold perfectly still through percussive material (validated: < 0.1 dB of per-hit motion on snare bursts) and glide only with the passage.
Gain staging follows the Omega philosophy: the leveler does the heavy lifting, so the multiband compressor only rides a gentle sweet spot instead of the 10–15 dB grind of conventional broadcast chains — factory presets are centered on a station-tuned AGGRESSIVE·LIGHT reference (bright, open-top FM voicing: near-flat leveler shape, light HM/HI compression, transients handed to the final limiters, loudness from the clippers) and calibrated against dense mastered music (a CD-loudness surrogate is part of the test suite) and land at ≈ 1.5–2 dB (LIGHT), ≈ 3–4 dB (MEDIUM) and ≈ 5 dB (HEAVY) of per-band GR, balanced across bands, with the leveler idling near centre instead of railing. The final WB/HF limiters catch only transients (mean well under 1 dB, peaks 3–4 dB) and loudness compensation comes from the baseband and composite clippers, not from compression depth. The WB limiter's release rides: brief peaks recover fast back to a slowly-moving platform gain instead of pumping to unity between words or beats (the validator proves 59 % of the working GR is held through gaps; after the platform unification it holds 89 %).
Latency of the whole DSP chain is ≈ 4.3 ms (plus the audio buffer).
No menus. Four SIGNATURE buttons (Aggressive / Smooth / Natural /
Custom) × three DENSITY buttons (Heavy / Medium / Light) = the 12
presets. LED bargraphs for every stage (input, AGC ride, per-band
leveler ±, per-band compressor GR, HF/WB limiting, output, MPX
modulation %), per-stage bypass, device/buffer selection, RDS panel,
and a TWEAK bar of thirteen knobs (drag ↑↓, scroll, double-click to
center). The ADVANCED button opens a second window with the deep
engineering controls — per-band compressor thresholds / ratios /
attacks / releases / makeup + knee + release platform, leveler band
targets, speed,
integration time and window, all AGC constants (incl. integration and
window), EQ frequencies and Qs, HF-limiter split & release, and
the WB limiter's ceiling / fast-release / platform time. "SAVE CUSTOM"
stores your tweaks into the CUSTOM signature; everything persists to
~/.omegafm/config.yaml.
BS.412 MPX power limiter — the European regulatory loudness governor, opt-in in the output section. Measures true composite power against the 19 kHz-deviation reference and rides only the audio part of the MPX (pilot and RDS injection are constitutionally untouched - verified to 0.000 dB), holding the honest sliding 60 s figure the regulator measures at the TARGET (default 0 dBr, with hysteresis so the gain glides instead of hunting). The readout lives next to the control and stays live even when limiting is off; enable and target persist as station setup like RDS. Note: enabling it costs loudness - that is its entire legal purpose.
8100-reference final section — the clip stage follows the classic Orban architecture: cubic soft clip, then the clipped-away difference below 1.7 kHz is added back phase-aligned (bass passes UNCLIPPED - fat, not crunchy), the 15 kHz brickwall, two passes of band-limited phase-aligned overshoot rounding, and a final stage that only ever touches sub-0.2 dB crumbs. Pre-emphasis carries the analog network's 20 kHz pole (without it the boost rises without limit - the classic 'harsh digital pre-emphasis' mistake). The HF limiter applies its gain as a dynamic 3.2 kHz shelf plus 15 % broadband, so it reads as program tilt, not a band being ducked. The phase rotator is four staggered first-order sections (120-210 Hz) with its own bypass in ADVANCED.
Performance — the gain-riding control loops JIT-compile through
Numba when installed (pip install numba - optional, pure-Python
fallback otherwise) and the garbage collector is parked while
streaming. Worst-case block time measured 15 ms against a 21 ms
budget with the full factory rack.
Distortion-controlled clipper — the clip stage is the big-iron kind: bass is pre-clipped off the rail on its own leg, the clip difference is split so bass/mid products (clean loudness) stay while HF products (audible sibilant 'spit') are suppressed, and the main clip runs with overshoot headroom under the 15 kHz brickwall so the post-filter safety never hard-shears the filter ring. Measured on a worst-case continuous-sibilance torture probe: odd-order IMD at the factory drive improved from a -24.6 dBc hard floor (unfixable by any clipper setting) to -27.4 dBc, reaching -43.7 at moderate drive, with loudness within 1 % and MPX peaks intact.
Power-on default — every launch boots into the station-tuned
factory default: the full nine-module rack (repair, enhancement and
governance, operator calibrated) plus the reference sound
— AGGRESSIVE · LIGHT (shipped as
presets/factory_default.json too), so the app always comes up in the
exact validated state. Only station identity and local setup persist
between sessions - RDS, injection levels, I/O trims, device and buffer
choices. Your sound lives in preset files: IMPORT one after boot.
Repair plugins (de-clipper, dehummer) include a TEST knob that
simulates their target damage so you can hear and meter them working
on any source - never leave TEST on for air.
Preset files — IMPORT / EXPORT buttons read and write portable JSON presets, applied live. Presets carry the plugin rack too: which plugins are enabled and every knob value — importing a preset reproduces the complete sound including its modules (plugins not named by the preset are disabled; files from before this feature leave the rack untouched). A preset is the sound only: the whole processing chain including bypasses, EQ, all dynamics constants and clipper drive. Station identity and local setup (RDS text/PI, pre- emphasis region, input gain, output level, pilot & RDS injection) are deliberately excluded so importing someone else's sound never clobbers your station. Files carry a format tag + version; on import unknown keys are ignored, missing keys fall back to defaults, every value is type-checked and clamped to safe DSP ranges — a hand-edited file can't blow up a filter. Round-trips are exact (validated), the exported DENSITY comes along but stays a live dimension afterward, and the imported preset's name lights up on the front-panel lamp.
Fully scalable canvas — the panel is drawn at a 1600×900 reference
inside a QGraphicsView that re-fits it (aspect-locked) on every
resize. On any window or screen size the entire interface simply
scales; nothing ever clips:
| 1600×900 | 880×540 |
|---|---|
|
 |
The validated core chain never needs rebuilding to grow: drop a single
.py file into plugins/ (or ~/.omegafm/plugins/), press
PLUGINS on the panel, tick it, and it is inserted live into the
running chain at the point it declares. A plugin file is ordinary
Python with a declarative manifest:
PLUGIN = {
"id": "stereo_widener", "name": "Stereo Widener", "version": "1.0",
"insert": "post_input", # post_input / post_agc / post_eq /
# post_bass / post_multiband (48 kHz stereo)
"params": [{"key": "width", "label": "WIDTH", "min": 0, "max": 2,
"default": 1.25, "fmt": "{:.2f}", "suffix": "x"}],
"meters": [{"key": "corr", "label": "CORR", "mode": "bipolar",
"lo": -1, "hi": 1}],
}
class Plugin:
def __init__(self, fs, channels=2): ...
def set_params(self, **kv): ...
def process(self, x): # float64 (N,2) in -> same shape out
return xThe CONTROLS button opens the plugin's own window - knobs and LED meters are auto-generated from the manifest and read the running instance live. Enable state and knob values persist across sessions.
The core is protected: every plugin call is sandboxed - on any exception or wrong output shape the plugin is instantly auto-bypassed with its traceback shown in the list, and audio continues untouched through the validated chain (this containment is itself a validation check). A working Stereo Widener ships as the reference example: mid/side WIDTH with a MONO< bass-mono guard and a live L/R correlation meter.
pip install -r requirements.txt
python run_omegafm.pyInput and output run as independent streams, so any input device works with any output device (different drivers / host APIs are fine), each side honours its own selection or its own system default, and the status bar shows the resolved device names while running. In MPX mode only the output runs at 192 kHz - the input stays at native 48 kHz. A small FIFO (2-block prefill) joins the two clocks; sustained clock drift between cheap devices is absorbed by drop-oldest and counted as an xrun.
Pick input/output devices, choose Stereo L/R @48k (normal listening — the monitor is de-emphasized so it sounds flat) or MPX composite @192k, press START.
- MPX mode needs a true 192 kHz interface whose output is flat to ~60 kHz, feeding an FM exciter's composite input. Consumer DACs often roll off above ~20 kHz and will kill the 57 kHz RDS subcarrier.
- On Windows, ASIO devices are tagged in the device list
(
pip install sounddeviceincludes ASIO support via PortAudio). - Buffer sizes: 256 / 512 / 1024 (default) / 2048.
python tools/validate_chain.py # 18 automated checks + MPX spectrum plot
QT_QPA_PLATFORM=offscreen python tools/render_ui.py # panel screenshotsThe validator proves: crossover flatness, pre-emphasis accuracy vs the analog curve, RDS CRC integrity, gain-riding/GR behaviour on program material, ≤ 100 % modulation, pilot/RDS placement and >45 dB out-of-band suppression in the Welch spectrum, that the WB/HF limiters only tickle while loudness is kept, that the WB release rides its platform instead of pumping word-to-word, that the leveler holds still through snare-like hits, DSP latency < 10 ms, and 75 dB decoded stereo separation.
omegafm/
dsp/filters.py biquads, LR4 crossover, pre-emphasis, FIR designs
dsp/dynamics.py look-ahead limiter, HF limiter, AGC/leveler, comp, clippers
dsp/resample.py polyphase 4x up/down converters
processing_chain.py FrontChain48 + FinalSection192
stereo_generator.py MPX generator + composite clipper + pilot/RDS injection
rds.py EN 50067 encoder
params.py defaults, signatures x densities, thread-safe store
processor.py sounddevice engine (stereo & MPX modes), Controller
meters.py, config.py, ui.py, main.py
tools/
validate_chain.py offline proof of the whole chain
render_ui.py offscreen screenshots (scaling demo)