What Is Audio Resampling?

Digital audio grew up with two incompatible standard rates, and their coexistence is the root cause of most resampling. Music inherited 44.1 kHz from the CD; video, film, and broadcast standardized on 48 kHz. Studios additionally record at 88.2, 96, or 192 kHz for production headroom. Whenever material crosses between these worlds, rates must be reconciled: a 48 kHz location recording destined for CD or streaming at 44.1 kHz; a 44.1 kHz music track imported into a 48 kHz video edit; a 96 kHz studio master downsampled for distribution; voice recordings at 16 or 22.05 kHz normalized into a modern project. Playback systems force the issue too — a sound card runs at one rate at a time, so everything else gets converted to match. Since the two rate families (44.1/88.2/176.4 and 48/96/192) do not divide into each other cleanly, crossing between them is the everyday case, not the exception, and doing it cleanly matters.

Conceptually, resampling reconstructs the continuous waveform the samples describe, then measures it again at the new rate. Downsampling (to a lower rate) has a mandatory first step: an anti-aliasing low-pass filter must remove all frequencies above the new, lower Nyquist limit — going 96 to 44.1 kHz means discarding content above 22.05 kHz — because anything left above that limit would fold back into the audible band as aliasing distortion. Then new sample values are computed at the new spacing. Upsampling (to a higher rate) interpolates: it calculates what the reconstructed waveform's value would be at the new, closer-spaced instants, using high-quality filtering (windowed-sinc interpolation in good resamplers). The crucial asymmetry: downsampling genuinely removes information (the frequencies above the new ceiling), while upsampling adds no information whatsoever — 44.1 kHz audio raised to 96 kHz contains exactly the same content, just described by more samples. Neither direction can ever recover or invent detail; resampling is re-measurement, not enhancement.

Some rate conversions are mathematically comfortable and some are not. Halving and doubling within a family — 96 to 48, 88.2 to 44.1 — involves a simple 2:1 ratio: the converter filters and takes every other sample, and even mediocre implementations do this well. Crossing between the families is the hard case: 44.1 and 48 kHz relate in the ratio 147:160, so no simple sample-keeping scheme exists — the converter must effectively reconstruct the waveform and re-measure it at entirely new time positions, demanding long, precise interpolation filters. Historically this is where cheap converters audibly failed; modern algorithms handle it transparently, but it remains the conversion where implementation quality shows. Two practical corollaries follow. First, when recording for a known destination, record in the destination's family (48 kHz for video work, 44.1 for CD-bound audio) so any later conversion is the easy kind. Second, when a 147:160 conversion is unavoidable, do it once, with a good resampler, at the final stage — not repeatedly along the way.

Upsampling is widely misunderstood as an upgrade. It is not: converting 44.1 kHz audio to 96 or 192 kHz creates a bigger file describing exactly the same sound — no new high frequencies, no added 'resolution,' nothing recovered. The information ceiling was fixed the moment the audio was first captured or encoded at its original rate. Legitimate reasons to upsample do exist, and they are all workflow reasons rather than quality reasons: matching a project or hardware rate that everything must share; feeding processors that operate better at high internal rates (many plugins internally oversample for exactly this reason); meeting a delivery spec that demands a particular rate. In those cases upsampling is harmless when done well — a clean interpolation adds essentially no artifacts. What it never does is make audio 'hi-res': a 44.1 kHz master sold or stored as 192 kHz is the sample-rate cousin of the fake FLAC — a bigger container around unchanged content. Upsample when a workflow requires it; never expecting improvement.

A well-implemented resampler is audibly transparent: in controlled blind tests, properly downsampled 44.1 kHz derived from a 96 kHz master is indistinguishable from the original on normal playback, because everything removed lies above human hearing. Poor resampling, by contrast, has recognizable failure modes. Aliasing — inadequate filtering lets ultrasonic content fold down into the audible band as inharmonic, metallic tones that were never in the music. Passband roll-off — an over-cautious filter dulls the top octave, audibly softening cymbals and air. Ringing — very steep filters can smear transients with pre- and post-echo. Phase distortion and imaging artifacts round out the list. The determining factor is the algorithm, not the rates: a high-quality converter crossing the awkward 147:160 ratio beats a sloppy one doing a trivial 2:1. This is why resampler quality comparisons (the well-known 'SRC comparison' plots of filter responses) exist, and why the fix for gritty-sounding rate conversion is almost always a better tool or setting, not a different target rate.

Resampling is often confused with two effects it deliberately avoids. Proper sample-rate conversion changes how the audio is described while preserving exactly how it sounds: same duration, same pitch, new rate. The confusion comes from the varispeed trick: if you take samples recorded at 44.1 kHz and simply play them back at 48 kHz without conversion, everything comes out roughly 9 percent fast and sharp — the tape-speed effect. That is mislabeled playback, not resampling, though it is exactly what goes wrong when a file's header says one rate and a system assumes another (the classic 'chipmunk audio' bug). Deliberate speed and pitch tools are different again: a speed changer alters tempo (using time-stretching to hold pitch constant, or resampling-plus-rate-relabeling for the tape effect), and a pitch changer shifts pitch while holding duration. Under the hood those tools may use resampling as one ingredient, but their goal is audible change, whereas resampling's goal is audible invariance. If a rate conversion changed the pitch or length, something used the wrong operation.

Most resampling in daily life happens silently. Operating-system mixers run all audio at one device rate — Windows shared mode, macOS Core Audio, and Android resample every stream that does not match, so 44.1 kHz music playing while the interface sits at 48 kHz is being converted live without notice. Bluetooth audio chains routinely resample to match codec and link rates. DAWs and video editors resample mismatched files on import or on the fly, at whatever quality their import converter offers. Sound cards resample to their clock; smart speakers and TVs resample to theirs. Streaming platforms resample uploads to their delivery standard. None of this is scandalous — modern system resamplers are decent — but it has two practical implications. First, obsessing over one pristine offline conversion is inconsistent if playback then passes through a mediocre live converter; second, for critical listening or mastering, matching the device rate to the material (or using exclusive/bit-perfect output modes) removes the hidden conversion entirely. Knowing where silent resampling lives explains many 'why does it sound different here' mysteries.

Rate conversion rarely travels alone — preparing a 24-bit/96 kHz master for CD means reaching 16-bit/44.1 kHz, which involves two distinct reductions with their own rules. Order and hygiene matter. Resample first at high bit depth (the converter's math benefits from the precision, and any intermediate rounding stays far below audibility), then reduce bit depth last, applying dither exactly once at that final truncation to convert quantization error into benign noise. Doing depth reduction before resampling, or dithering multiple times, adds avoidable noise; skipping dither entirely risks audible grain on quiet material. Inside DAWs this is mostly handled for you — internal processing runs at 32-bit float and export dialogs apply dither on the way out — but standalone conversions replicate the same sequence: convert rate, then depth, then dither, then encode. If the destination is a lossy format, the encoder handles final precision itself, so the practical chain becomes: resample to the target rate if needed, then feed the clean high-depth PCM straight to the MP3/AAC/Opus encoder.

The best resampling is none. If source and destination already share a rate, ensure no conversion sneaks in. Choose your rate at the start of a project and record in it: 48 kHz for anything destined for video, 44.1 kHz for CD-bound music, either for streaming (platforms accept both). Keep a whole project at one rate rather than mixing files that force per-import conversions. Avoid repeated resampling especially — like lossy transcoding, each pass can add tiny artifacts, and round trips (44.1 to 48 to 44.1) accumulate filter effects for zero benefit; when a conversion is needed, do it once, at the final stage, from the best source. Do not upsample seeking quality, and do not down-then-up-sample as part of casual editing. Archive masters at their native rate so future deliverables each get a single fresh conversion. In short: pick the right rate early, convert late, convert once — the transparent resampler is best kept idle.

Quality resampling is widely available, much of it free. SoX and its extracted soxr library are the open-source reference — the command-line 'rate' effect with high-quality settings is effectively transparent, and soxr powers resampling inside many other tools. FFmpeg performs resampling automatically whenever an output format or explicit flag requires it, and can be built to use soxr for its best quality. Audacity resamples cleanly on import and on project-rate export. Every major DAW — Pro Tools, Logic, Ableton, Reaper (whose resampler is notably configurable) — converts on import and bounce at professional quality. Dedicated mastering-grade converters (iZotope RX, Weiss Saracon) serve the most demanding workflows, though the audible gap between 'good' and 'best' has essentially closed. AudioUtils handles resampling automatically inside its browser-based conversions: when a target format or setting requires a different sample rate, the FFmpeg WebAssembly engine performs the conversion locally — no upload, no signup, and your files never leave your device.