Audio samples belong to a clock, not merely to a numbered array

Record the capture device's nominal rate, channel count, sample format, and timestamps tied to a defined clock. Count sample frames separately from interleaved scalar values: one stereo frame contains two channel samples but represents one sampling instant. A consumer that advances at exactly the nominal rate can slowly diverge from a producer whose oscillator is fast or slow. Network arrival variation is another effect; neither a packet timestamp nor a large queue alone tells whether drift came from the source clock, the receiver clock, or scheduling delays.

Clock error is often expressed in parts per million. Multiplying the fractional error by nominal samples per second gives an approximate sample imbalance per second when the consumer follows the nominal clock. Over long sessions, even a small error can cause repeated overflows or underflows if no correction exists. Use timestamp and occupancy trends to estimate rate mismatch, then apply a controlled correction such as asynchronous resampling where appropriate. Arbitrarily deleting individual samples can produce clicks and timing discontinuities instead of preserving a smooth waveform.

A rational converter raises the internal sampling grid, applies an appropriate low-pass filter, and selects output samples on the new grid. Downsampling needs suppression of frequencies that would fold into the output band; simply keeping every third sample preserves neither arbitrary high-frequency content nor a clean speech band. Real filters need a transition band, so the useful passband is not an ideal brick wall extending exactly to the new Nyquist frequency. Specify passband ripple, stopband attenuation, and latency for the intended application rather than calling every converter equivalent.

SciPy 1.16.0 documents resample_poly as polyphase conversion with a zero-phase low-pass FIR and explicit boundary-extension choices. That offline interface can center its result and use samples on both sides of an output position. A live causal converter must retain filter history and fractional phase across chunks, and it incurs a delay or a different boundary policy. Calling an offline routine independently on each packet with zero padding invents artificial boundaries. Preserve the continuous stream's conversion state instead of treating transport packet edges as signal endpoints.

Test long runs with known oscillator mismatch, impulses, tones near the band edge, and varying packet sizes. Compare total emitted sample counts, timestamp mapping, boundary transients, and spectral distortion with a continuous reference. Account for startup and final filter delay rather than silently trimming a different amount on every chunk. Audio format conversion also needs explicit clipping, channel mixing, and integer-to-floating scaling rules. A model that expects mono sixteen-kilohertz input should receive that representation through a verified converter, not through a renamed buffer whose actual clock and layout remain unchanged.

The waves lab illustrates only a consequence of phase mismatch between equal-frequency signals. Reset, explicitly set BOTH phase=0 degrees and amplitude ratio=1, predict resultant amplitude two, then Play or Step. Reset and re-set BOTH phase=180 degrees and ratio=1; predict zero. Edits retain t. This model does not advance independent sample clocks, filter a stream, resample data, or estimate drift. The controlled phase comparison motivates timing care, while the sample-count and rate calculations provide the actual clock accounting for this lesson.