Fractional delay (FD) filters are fundamental in digital signal processing when signals must be delayed by a non-integer number of samples with high accuracy and low distortion. Applications include sample rate conversion, beamforming, synchronization, and channel modeling. Among the various FD filter design methods, the Farrow structure has emerged as a particularly efficient and flexible approach. This presentation introduces the principles, implementation, and advantages of fractional delay Farrow filters.

We begin by reviewing the concept of FD filtering and the challenges associated with achieving arbitrary delays in discrete-time systems. Next, the Farrow filter structure is examined in detail, highlighting its polynomial-based decomposition and modular architecture that enables real-time tunability of the delay without redesigning the filter. The computational benefits of the Farrow approach, including its suitability for hardware implementation and low memory footprint, are discussed.

Simulation examples demonstrate the performance of different polynomial approximations in terms of delay accuracy, group delay distortion, and numerical stability. Finally, we outline practical applications where fractional delay Farrow filters offer significant performance gains. Attendees will gain a clear understanding of both the theoretical foundations and practical aspects of these filters, equipping them to apply Farrow structures effectively in DSP systems.

This guide was created with the help of AI, based on the presentation's transcript. Its goal is to give you useful context and background so you can get the most out of the session.

What this presentation is about and why it matters

This talk introduces fractional delay (FD) Farrow filters: a compact, tunable way to delay a discrete-time signal by a non-integer number of samples. Fractional delays are a practical necessity in many real systems where the desired timing does not align with a single sample period. Examples include sample-rate conversion, beamforming (where precise timing steers the array), timing and carrier synchronization in communications, channel modeling, and audio effects. The Farrow structure stands out because it implements a whole family of fractional delays with a single fixed filter bank plus a small polynomial-combiner, allowing real-time adjustment of the fractional delay parameter without redesigning filter coefficients.

Who will benefit the most from this presentation

This presentation is most useful for:

• DSP engineers who implement timing alignment, resampling, or multichannel synchronization.
• Students learning interpolation-based filter design and practical trade-offs between accuracy, complexity, and bandwidth.
• Hardware engineers targeting low-memory or low-latency implementations (FPGA/ASIC) where a reusable structure is preferable to many separate FIR filters.
• Audio and communications engineers who need variable, low-distortion delay with good phase behavior in a target frequency band.

What you need to know

Before watching, refresh the following concepts so you can follow the motivations and design choices in the talk:

• Sampling and discrete-time delay: delaying by an integer number of samples is trivial (index shift). Fractional delay means producing the signal value at non-integer sample offsets.
• Interpolation basics: methods such as Lagrange polynomial interpolation, sinc (ideal) interpolation, and cubic Hermite splines are common ways to approximate values between samples. Understand their trade-offs: order (filter length), passband accuracy, and behavior near Nyquist.
• FIR vs. IIR: most interpolation/Farrow implementations use FIR filters because they can be designed to be linear phase and stable; IIRs are possible but bring phase nonlinearity and stability concerns.
• Polynomial approximation: Farrow approximates a family of impulse responses as polynomials in the fractional parameter $\mu$. For example, the Farrow output takes the form $y[n]=\sum_k h_k(\mu)\,y_k[n]$, where each $y_k[n]$ is the output of a fixed subfilter and the $h_k(\mu)$ are polynomials in $\mu$.
• Horner's method: an efficient way to evaluate polynomials in $\mu$ using nested multiply–add operations, which is commonly used in Farrow combiners for low computational cost.
• Integer vs fractional separation: practical designs usually separate integer delay (simple buffer) from the fractional part handled by the Farrow filter. This reduces the polynomial range and keeps subfilters shorter.
• Performance metrics: be ready to interpret group delay, phase linearity, frequency-dependent error (especially near Nyquist), gain variation vs. $\mu$, and numerical issues in fixed-point implementations.

Glossary

• Fractional delay (FD): delaying a discrete-time signal by a non-integer number of samples.
• Farrow structure: a filter architecture that uses a bank of fixed subfilters whose outputs are combined by polynomials in a fractional parameter $\mu$ to realize variable FD.
• Subfilter (or basis filter): a fixed FIR filter in the Farrow bank; each produces a different delayed response that is linearly combined.
• $\mu$ (fractional parameter): the normalized fractional delay (typically in [0,1) for the fractional part) that controls the polynomial weights in the combiner.
• Lagrange interpolation: polynomial interpolation that exactly fits a set of sample points; commonly used as a basis for Farrow coefficient design.
• Cubic Hermite spline: an interpolator that enforces continuity of value and first derivative (C1), often yielding smoother behavior than simple polynomials.
• Group delay: the frequency-dependent delay experienced by different spectral components; linear group delay implies no phase distortion across the band.
• Horner's method: an efficient algorithm for polynomial evaluation using nested multiplications and additions.
• Polyphase filters: an alternative resampling approach that can be more efficient than Farrow for fixed-rate converters or large integer changes.
• Nyquist frequency: half the sampling rate; interpolation accuracy typically degrades as components approach Nyquist, requiring higher-order designs for wideband FD.

Tip: While watching, focus on the interplay between filter order, available bandwidth, computational cost, and the decision to separate integer and fractional delays. The demo and provided Python code will make these trade-offs concrete, so downloading the example code after the talk is a practical next step.