Audio Effects Pipeline | DSP-101

The Complete Signal Chain

A professional audio processor is not a single algorithm — it is a pipeline of specialized stages, each solving a different problem with a different DSP technique. An equalizer shapes the frequency balance. A reverb unit adds the acoustic impression of a room. A compressor tames dynamic swings. A delay or chorus effect sculpts the stereo image and adds richness. Together these stages form the audio effects pipeline that sits between a microphone and a speaker in every mixing console, DAW, and live sound system on the planet.

Understanding this pipeline is understanding applied DSP. Every concept from earlier modules — FIR and IIR filters, the DFT, convolution, feedback systems — reappears here in a concrete, audible form. This lesson walks through each stage: what it does acoustically, how it is implemented in DSP, and how the stages connect into a coherent chain.

Core Idea

Each stage in the audio pipeline is independently designed, then connected via a simple wet/dry mix control (how much of the processed signal to blend with the original). This modularity means you can swap, reorder, or bypass individual effects without redesigning the whole chain — the same principle used in every DAW plugin architecture.

The Parametric Equalizer

An equalizer adjusts the amplitude of different frequency bands, shaping the tonal character of the audio. A parametric equalizer gives the engineer three controls per band: center frequency (where the boost or cut is applied), gain (how many dB to add or subtract), and bandwidth Q (how wide the affected band is). These three parameters together can sculpt any smooth frequency response imaginable.

The building block of any parametric EQ is the second-order IIR filter, commonly called a biquad. Each biquad implements one EQ band with just five coefficients and a two-pole, two-zero transfer function. Because a biquad is IIR, it achieves complex frequency shaping with a tiny computational footprint — just five multiplications and four additions per sample, regardless of how narrow the band is.

Biquad Transfer Function

H(z) = \frac{b_0 + b_1 z^{-1} + b_2 z^{-2}}{1 + a_1 z^{-1} + a_2 z^{-2}}

The biquad filter is described by five coefficients: b₀, b₁, b₂ (numerator/zeros) and a₁, a₂ (denominator/poles). Audio libraries like RBJ (Robert Bristow-Johnson) Audio EQ Cookbook provide closed-form formulas for computing these coefficients from the user-facing parameters (frequency, gain, Q).

A practical parametric EQ chains multiple biquads in series. A four-band EQ uses four biquads, consuming 20 multiplications and 16 additions per sample — a trivial load on any modern DSP. The four standard band types are: low shelf (boost/cut all frequencies below a corner), peaking bell (boost/cut around a center frequency), notch (deep cut at a specific frequency, for removing hum or resonances), and high shelf (boost/cut all frequencies above a corner).

Coefficients per biquad EQ band

9 MAC

Operations per sample per biquad band

±18 dB

Typical parametric EQ boost/cut range

Reverb: Room in a Box

Reverb simulates the acoustic behavior of a physical space — the way sound reflects off walls, ceiling, and floor before decaying into silence. A dry (close-miked) recording sounds clinical and unnatural without reverb; adding it places the sound in a perceivable acoustic environment, whether a living room, concert hall, or cathedral.

Algorithmic reverb uses a network of feedback delay lines, comb filters, and all-pass filters to create dense, diffuse reflections. The classic Schroeder and Moorer designs chain a bank of parallel comb filters (which produce periodic echoes at different delays) into a series of all-pass filters (which diffuse the echoes without changing their power). The result is an exponentially decaying tail whose character depends on the chosen delay lengths and feedback gains. Algorithmic reverb is computationally cheap and easy to control but can sound metallic or colored at extreme settings.

Convolution reverb takes a different approach: measure the actual impulse response of a real room (by recording the room’s response to a starter pistol or a swept sine) and then convolve any audio signal with that impulse response. The result is mathematically perfect — the dry signal sounds as if it were recorded in that exact room. The computational cost is the price: convolving with a two-second reverb tail at 48 kHz requires a 96,000-sample convolution. Overlap-add or overlap-save partitioned convolution reduces this to a tractable sequence of short FFTs.

Convolution Reverb

y[n] = \sum_{k=0}^{L-1} x[k] \cdot h[n-k]

The reverberated output y[n] is the convolution of the dry input x[n] with the room impulse response h[n]. In practice, h[n] may be tens of thousands of samples long, so partitioned convolution splits it into short blocks and uses the FFT-based fast convolution (overlap-add) to achieve O(N log N) complexity.

Dynamic Range Compression

A compressor reduces the dynamic range of a signal — the difference between its loudest and quietest moments. Drums can have peaks 30 dB louder than their average level; without compression, either the peaks clip the output or the average level is too quiet to be heard clearly. A compressor automatically turns down the loud parts, allowing the overall level to be raised until the quiet parts fill the mix.

A compressor is defined by five parameters: threshold (the level above which gain reduction begins), ratio (how aggressively levels above the threshold are reduced; a 4:1 ratio means 4 dB in yields 1 dB out above threshold), attack time (how quickly the gain reduction is applied after the signal crosses the threshold), release time (how quickly it recovers), and makeup gain (applied after compression to restore the overall level).

Gain Computer

y_{\text{dB}} = \begin{cases} x_{\text{dB}} & x_{\text{dB}} \leq T \\ T + \dfrac{x_{\text{dB}} - T}{R} & x_{\text{dB}} > T \end{cases}

The gain computer maps input level x_dB to output level y_dB. Above the threshold T, the gain is reduced by a factor of (1 − 1/R), where R is the compression ratio. Below threshold, the signal passes unmodified. The gain reduction signal is then smoothed by the attack/release envelope follower before being applied to the audio.

The DSP implementation of a compressor involves two loops: a level detector (an envelope follower that tracks the signal’s RMS or peak level with separate attack and release time constants) and a gain computer (which maps the detected level to a gain reduction amount via the threshold/ratio law). The computed gain is converted to linear scale and multiplied sample-by-sample against the audio. A limiter is simply a compressor with a very high ratio (10:1 or ∞:1), used as a hard ceiling to prevent any sample from exceeding 0 dBFS.

Look-Ahead Limiting

A zero-attack limiter would clip the first sample of each transient before the gain computer can react. Look-ahead solves this by buffering the audio 5–10 ms into the future, giving the gain computer advance notice to begin gain reduction before the transient arrives. This adds a fixed latency but produces a transparent, distortion-free limit.

Delay and Chorus Effects

A delay effect plays back one or more echoes of the input at fixed time intervals. The delay time, feedback level (how much of each echo feeds back into the delay line), and mix determine the character — from a subtle slap-back (100–150 ms, no feedback) to cascading, rhythmic repeats. In DSP, a delay is simply a circular buffer read from an offset position relative to the write head: output = buffer[write_ptr − delay_samples]. No computation beyond a buffer read and a feedback multiply is needed.

A chorus effect creates the sensation of multiple instruments or voices playing together, thickening the sound. It works by modulating the delay time of one or more delay lines with a low-frequency oscillator (LFO), typically a sine wave at 0.5–5 Hz with a depth of 5–25 ms. The slight pitch variation introduced by the changing delay time mimics the natural detuning between ensemble performers. Reading from a non-integer delay position requires interpolation — linear or all-pass interpolation — to avoid aliasing artifacts from the fractional sample delay.

Modulated Delay (Chorus)

d(n) = D + A \sin\!\left(2\pi f_{\text{LFO}}\, n / f_s\right)

The chorus delay time d(n) is the sum of a base delay D and a sinusoidal modulation with depth A and LFO frequency f_LFO. The instantaneous pitch shift is proportional to the rate of change of d(n): Δf = −f₀ · (dd/dn) / f_s, which oscillates between −A·2πf_LFO/f_s · f₀ and +A·2πf_LFO/f_s · f₀.

Assembling the Complete Chain

The order of stages in an audio pipeline matters acoustically. The conventional ordering for most production scenarios is:

Input

Gain & trim

→

Tonal shaping

→

Comp

Dynamics control

→

Reverb

Space & depth

→

Delay

Echo & chorus

→

Limiter

Output ceiling

EQ before compression ensures the compressor reacts to the tonally corrected signal, not artifacts or problem frequencies. Reverb after compression prevents the compressor from pumping on the reverb tail. The final limiter acts as a safety net, catching any inter-sample peaks that exceed 0 dBFS before digital-to-analog conversion.

Each stage exposes a wet/dry mix parameter (from 0% dry to 100% wet). Parallel processing — running the dry signal and the effect simultaneously and blending the output — is standard for reverb and delay, preserving the attack transients that would be smeared by the effect alone. Some compressors also offer a parallel compression (New York compression) mode for the same reason.

Latency Budget

Each stage adds processing latency. Biquad EQ and compressors are sample-by-sample and add zero or single-sample latency. Block-based convolution reverb adds one block duration (typically 512–2048 samples = 10–43 ms at 48 kHz). Look-ahead limiting adds its look-ahead window (5–10 ms). The total pipeline latency must remain below the threshold of perceptibility (<20 ms for live monitoring) or be compensated by the host DAW in recording contexts.

Next lesson: DSP in Communications — OFDM receivers, channel equalization, timing synchronization, and error correction decoding — the DSP engine inside every Wi-Fi and 4G/5G modem.

Key Takeaways

A parametric EQ is a chain of biquad IIR filters, each with just five coefficients and nine MAC operations per sample, capable of implementing low shelf, peaking, notch, and high shelf bands.
Algorithmic reverb uses networks of comb and all-pass filters to create dense, diffuse decay tails; convolution reverb convolves the input with a real room's impulse response for perfect acoustic realism at higher computational cost.
A compressor consists of a level detector (envelope follower) and a gain computer (threshold/ratio law); a limiter is a high-ratio compressor with look-ahead to catch transients before they clip.
Delay effects use circular buffers; chorus modulates the delay time with an LFO, requiring fractional-sample interpolation to avoid aliasing from the non-integer delay position.
The canonical pipeline order — EQ → Compression → Reverb → Delay → Limiter — reflects acoustic logic: EQ corrects tone before the compressor reacts to it, and the limiter provides a final safety ceiling.
Latency accumulates through the pipeline; convolution reverb and look-ahead limiting are the largest contributors and must be compensated in low-latency live applications.