Theory of operation

Signal generator - the stimulus source. Produces calibrated test signals: a spectrally pure sine for distortion measurements, two tones for intermodulation, white/pink noise, linear and logarithmic sweeps, and a compensated sine that actively cancels the playback chain's own distortion so the DUT receives a cleaner signal than the DAC alone can produce. Output level is set in volts, against the DAC calibration.

Oscilloscope - the time-domain view. This is where you look first: is the signal there at all, is it clipping, does it carry DC or mains hum, what are V_pp / V_rms / frequency? Triggering, per-channel scaling and a running measurement table work like on a bench scope. A spectral number is only worth trusting after the time-domain picture looks sane.

FFT analyser - the frequency-domain workhorse. Shows the spectrum and measures THD, THD+N, SNR, SINAD/ENOB and the individual harmonics; in two-tone mode the intermodulation products. Coherent averaging digs signals out far below the single-shot noise floor; mains rejection, calibration de-embedding and a manual fundamental reference handle real-world measurement setups (notch filters, attenuators, known source levels).

Frequency response - the swept measurement. Plays a logarithmic sweep through the DUT and computes its transfer function - magnitude and phase across the whole band in seconds. Results can be compared against RIAA / IEC curves or saved as calibration files that the FFT analyser later de-embeds from its own measurements.

Underneath all of them sit the two infrastructure layers this chapter describes: the audio backend (the bit-exact bridge to the sound hardware) and the ring buffer (the shared memory that lets every view read the same capture independently).

• Coherent averaging digs below the single-shot floor - and noise sets the limit. Every deep measurement (the harmonics, the two-tone products, the pre-distortion correction) sums phase-aligned blocks so the tone holds while the noise falls as 1/√n. The floor you reach is set by the bin signal-to-noise ratio and the effective number of averages - not by any trick.
• Measure the reference afresh, never integrate it. Phase references and frequency locks - the per-block de-rotation angle, the frequency-lock loop, the manual fundamental - are read on each block, never carried by a free-running counter. A measured reference averages out; an integrated one random-walks and slowly de-coheres the tone.
• Drift is a walking phase, not a lost level. Converter clock mismatch and notch-component temperature drift show up as a slow phase walk, which the analysis strips block by block. Tracked as phase, they cost coherence, not magnitude - which is why the wide-band level numbers stay solid while the fundamental sits in a drifting null.
• Exact frequencies by integer phase. The DDS accumulator, every harmonic correction tone and the sweep reference are formed by integer multiplication of one master phase, so a tone lands exactly at h·f and never beats against the distortion it measures or cancels, however long it runs.
• Windows are chosen for side-lobe margin, not noise. Above each window's recommended overlap every window reaches the same floor in the same wall-clock time, so the choice (HFT248D at 87.5 %) is about keeping a strong tone's leakage off the faint bins, not about noise.
• Calibration at the edges, normalised in the middle. ADC and DAC full-scale voltages are applied only at display and output; the signal data stays in normalised units and full precision throughout, so references and exports are the mathematically exact waveform.