Skip to content

FilterCompressor-Teletronix LA-2A.md

Latest commit

 

History

History
208 lines (143 loc) · 9.6 KB

FilterCompressor-Teletronix LA-2A.md

File metadata and controls

208 lines (143 loc) · 9.6 KB

Teletronix LA-2A Leveling Amplifier

Emulating the optical compressor that defined the sound of recorded voice for six decades

The Legend of the LA-2A

In 1965, Jim Lawrence founded Teletronix Engineering and introduced a compressor that would become synonymous with vocal recording: the LA-2A Leveling Amplifier. Decades later, engineers still reach for LA-2As—original units now command five-figure prices—whenever a voice needs warmth without aggression, control without constraint.

The LA-2A's magic lies in its T4 electro-optical attenuator: an electroluminescent panel paired with a cadmium sulfide photoresistor. Light output changes with the input signal; photoresistor conductance follows the light. This indirect coupling creates inherently smooth gain changes that electronic circuits struggle to replicate.

Bill Putnam, founder of Universal Audio and the man who acquired Teletronix in the 1960s, described the LA-2A's character simply: it "treats your signal lovingly."

What Makes the T4 Cell Special

The T4's response isn't linear or instantaneous—and that's the point. The photoresistor's conductance changes occur on a molecular level, creating attack and release curves that no capacitor-based design can match:

Characteristic T4 Behaviour Sonic Result
Attack ~10ms fixed Transients pass through naturally
Release (initial) 60ms to 50% Quick recovery from peaks
Release (full) 1–15 seconds Graceful return, no pumping
Ratio Programme-dependent Harder compression on louder signals
Knee Inherently soft Gradual onset, musical transition

The two-stage release is the LA-2A's signature: short peaks recover quickly, sustained compression releases slowly. This creates the "levelling" effect—consistent output without the mechanical feel of VCA compression.

Jivetalking's Implementation

Jivetalking captures the LA-2A's programme-dependent behaviour through FFmpeg's acompressor filter with spectral-adaptive parameter tuning. Pass 1 measurements drive every parameter, approximating the way the T4 cell naturally responds to different programme material.

Design Philosophy

T4 Characteristic Implementation Strategy
Fixed 10ms attack Base 10ms with ±2ms transient adaptation
Two-stage release LRA + spectral flux determine release time
Programme-dependent ratio Spectral kurtosis modulates 2.5–3.5:1 range; high-crest override can push toward 5.0:1 based on deficit severity
Soft knee Centroid-adaptive knee (3.5–5.0)
Warmth on dark voices Skewness detection triggers knee + release boost

Adaptive Parameter Tuning

High-Crest Override

When a recording's predicted limiter ceiling deficit is positive, the sub-tuners alone cannot prevent Pass 4's alimiter from clamping. The high-crest override detects this condition using Pass 1 measurements and applies proportional floor values that sub-tuners cannot reduce below.

Detection mirrors calculateLimiterCeiling() in normalise.go:

gainRequired = NormTargetLUFS - InputI
projectedTP  = InputTP + gainRequired
idealCeiling = LoudnormTargetTP - gainRequired - safetyMargin
deficit      = minLimiterCeilingDB - idealCeiling

When deficit > 0, severity scales linearly: clamp(deficit / 6.0, 0.0, 1.0).

Override floors (linear interpolation from severity 0.0 to 1.0):

Parameter Default At severity 1.0 Formula Enforcement
Threshold -18 dB -40 dB lerp(-18.0, -40.0, severity) min(tunerResult, floor)
Ratio 3.0 5.0 lerp(3.0, 5.0, severity) max(tunerResult, floor)
Release 200 ms 350 ms lerp(200.0, 350.0, severity) max(tunerResult, floor)
Knee 4.0 6.0 lerp(4.0, 6.0, severity) max(tunerResult, floor)

The override blends from no change (deficit just above 0) to maximum aggressiveness (deficit >= 6 dB). Sub-tuners may push values beyond the floor but cannot pull them back.

Attack: Preserving Consonant "Pluck"

The LA-2A's fixed 10ms attack is slow enough to let word onsets through—critical for podcast intelligibility. We honour this baseline with minimal adaptation for extreme cases:

Transient Character MaxDifference Attack
Sharp transients > 25% 8 ms
Normal speech 10–25% 10 ms
Soft delivery < 10% 12 ms

Release: Approximating Two-Stage Recovery

The T4 cell's two-stage release gives the LA-2A its breathing room. We approximate this by scaling release time with loudness range and spectral flux:

Source Character LRA Spectral Flux Release
Expressive speech > 14 LU > 0.025 300 ms
Standard podcast 8–14 LU 0.008–0.025 200 ms
Compressed delivery < 8 LU < 0.008 150 ms

Warm voice boost: Dark voices (skewness > 1.5) add 30ms release to preserve body. Heavy compression boost: Large LUFS gap (>15dB) adds 50ms, mimicking the T4's slower recovery after sustained compression. High-crest override: When active, extends the release range to 350 ms to accommodate recordings with large predicted ceiling deficits.

Ratio: The Levelling Character

Real LA-2As don't have a fixed ratio—the T4 cell compresses harder on louder signals. We approximate this programme-dependency using spectral kurtosis (peaked vs. flat spectrum):

Spectral Character Kurtosis Ratio
Peaked/tonal harmonics > 10 2.5:1
Standard speech 5–10 3.0:1
Flat/noise-like < 5 3.5:1

Very wide dynamic range (>35dB) adds +0.5 to the ratio for extra control.

Threshold: Signal-Relative

The LA-2A's "Peak Reduction" knob sets threshold relative to the incoming signal. We calculate threshold from peak level minus a dynamic range-dependent headroom:

Dynamic Range Headroom Compression Depth
> 30 dB 20 dB Heavy levelling
20–30 dB 15 dB Standard LA-2A
< 20 dB 10 dB Light levelling

Threshold range: -40 dB to -12 dB (clamped for safety). The high-crest override can push threshold to -40 dB, beyond the range the normal tuner typically selects.

Knee: T4 Softness

The T4 cell's gradual gain changes produce an inherently soft knee. We adapt knee softness based on voice character:

Voice Character Spectral Centroid Knee
Dark/warm voice < 4000 Hz 5.0
Normal voice 4000–6000 Hz 4.0
Bright voice > 6000 Hz 3.5

Warm voices (skewness > 1.5) add +0.5 knee for extra softness.

Mix: Honouring 100% Wet

Real LA-2As have no parallel compression—they're 100% wet. We default to full wet, reducing mix only for problematic recordings where dry signal masks artefacts:

Recording Quality Noise Floor Mix
Studio clean < −65 dBFS 1.0
Home office −65 to −45 dBFS 0.93
Noisy environment > −45 dBFS 0.85

Makeup Gain: Conservative Compensation

Makeup gain compensates for compression but shouldn't over-drive downstream processing:

overshoot = peak_level − threshold
reduction = overshoot × (1 − 1/ratio)
makeup = reduction × 0.65

Clamped to 1–5 dB. Let normalisation handle the rest.

Configuration

FilterChainConfig Fields

Field Type Range Default Purpose
LA2AEnabled bool — true Enable/disable filter
LA2AThreshold float64 −40 to −12 dB −24 dB Compression threshold
LA2ARatio float64 2.0–5.0 3.0 Compression ratio
LA2AAttack float64 8–12 ms 10 ms Attack time
LA2ARelease float64 150–350 ms 200 ms Release time
LA2AMakeup float64 1–5 dB 2 dB Makeup gain
LA2AKnee float64 3.5–6.0 4.0 Knee softness
LA2AMix float64 0.85–1.0 1.0 Wet/dry mix
LA2AHighCrestActive bool — false Whether high-crest overrides were applied
LA2AHighCrestDeficit float64 dB 0.0 Predicted ceiling deficit
LA2AHighCrestSeverity float64 0.0–1.0 0.0 Override scaling factor
LA2AHighCrestProjectedTP float64 dBTP 0.0 Projected true peak before compression

FFmpeg Filter Specification

acompressor=threshold=-24dB:ratio=3:attack=10:release=200:makeup=2dB:knee=4:mix=1

Pipeline Integration

DS201 Gate (dynamics cleanup) → Dolby SR (noise polish) → LA-2A (levelling) → Limiter

Division of responsibility:

  • DS201 Gate: Inter-phrase silence cleanup
  • Dolby SR: Spectral noise floor polishing
  • LA-2A: Dynamic range levelling with warmth
  • Limiter: Final peak control

The LA-2A operates on already-cleaned audio, so it can focus on its strength: gentle, programme-dependent levelling that evens out volume variations while preserving the natural character of speech.

References