SDR in Pure Go, Part 5: Tuning & Channelization — DDC vs. Polyphase

By

Part 5 of SDR Internals. One dongle can watch megahertz of spectrum at once. This post is about pulling many narrow channels out of that wide capture — and the two strategies GopherTrunk picks between at runtime.

In this post

  • Why “one dongle, many channels” matters for a trunking scanner.
  • Two approaches: a per-channel digital down-converter (DDC) versus a shared polyphase channelizer.
  • The Strategy pattern in Go: a Bank interface with two implementations, chosen at runtime.

What channelization does

A trunked radio system spreads its control channel and voice channels across a band. With a wideband SDR you can capture the whole band once and then tune inside the samples to as many channels as you want — no extra hardware. That “tune in software” step is a digital down-converter: shift the channel to baseband with an NCO, low-pass filter it, and decimate to a manageable rate (~48 kHz, the symbol clock for 4800-baud protocols).

There are two ways to do this for many channels at once, and they have opposite cost profiles:

  • DDC bank — one NCO + resampler per channel. Cost grows linearly with the number of taps, but the channels can sit at any offset.
  • Polyphase channelizer — a single filter bank splits the whole band into M evenly-spaced bins in one shared operation. Cheap per channel, but the channels must lie on the bin grid. (reference)

How GopherTrunk implements it in Go

internal/dsp/tuner defines the abstraction and internal/dsp/channelizer provides the heavy machinery. The key is a single interface — a Bank — with two interchangeable implementations:

// internal/dsp/tuner — the strategy contract (shape)
type Bank interface {
    // Process consumes one wideband chunk and writes the
    // narrowband output for every configured tap.
    Process(wide []complex64) [][]complex64
    Taps() []float64 // tuned offsets, in Hz
}
  • DDCBank instantiates one NCO + polyphase resampler per requested offset. Linear cost, arbitrary spacing — ideal when you only need a handful of taps at awkward frequencies.
  • ChannelizerBank runs the wideband stream through one polyphase decomposition + FFT rotation, emitting one sample per channel for every M input samples, then fine-tunes each bin with a small DDC. Shared cost, grid-aligned — ideal when you want many evenly-spaced channels (for example, a P25 Phase 1 control channel and its voice channels on one dongle).

The caller — the wideband voice tuner in internal/sdr/wbvoice, or the control decoder — picks the implementation based on how many channels it needs and how they’re spaced, then uses only the Bank interface.

The design principle: the Strategy pattern

DDC-per-channel and the polyphase channelizer solve the same problem with different trade-offs. That is the definition of the Strategy pattern: encapsulate interchangeable algorithms behind a common interface and choose between them at runtime.

How that principle shaped the Go code

  • One interface, two algorithms. Both banks satisfy Bank, so the surrounding code is written once against the interface. Swapping strategies is a constructor choice, not a rewrite.
  • The decision is data-driven. Number of taps and their spacing — not a compile-time flag — decide which bank to build. Even channel spacing favors the channelizer; sparse, irregular taps favor the DDC bank.
  • Go interfaces keep it implicit. Neither bank declares “I implement Bank”; they just have the right methods. That structural typing makes it trivial to add a third strategy later (say, an FFT-overlap-save filterbank) without touching the callers.
  • Composition over inheritance. Each bank is built from the Part 4 primitives — NCOs, resamplers, polyphase filters — rather than extending a base class. The strategy is assembled, not inherited.

Where this goes next

Channelization is what makes GopherTrunk’s “monitor a whole system on one SDR” feature possible. A future deep dive can compare DDC and channelizer CPU cost as the tap count climbs, and explain the FFT-rotation math at the channelizer’s core. Next up, the channels we just carved out get demodulated.

FAQ

When should I use a channelizer instead of per-channel DDCs? When you need many channels and they fall on a regular grid (most trunked systems do). The channelizer amortizes one big filter across all of them. For a few channels at irregular offsets, separate DDCs are simpler and just as fast.

Does channelization need a wideband SDR? It needs the SDR’s sample rate to cover all the channels you want at once. An RTL-SDR at 2.4 MS/s spans ~2 MHz of usable bandwidth — enough for many trunked systems’ control-plus-voice spread.

Why decimate to ~48 kHz? Most digital voice protocols run at 4800 symbols/s; ~48 kHz gives a clean integer oversampling ratio for the matched filter and timing recovery stages that follow.

Series navigation

Part 5 of 14 · ← Part 4 · Next → Part 6: Demodulation

#sdr#go#dsp#channelizer#ddc#software-design