What Is Lossy Audio Compression?

Lossy codecs exploit the difference between what a microphone captures and what a human actually perceives. Raw PCM audio stores every detail of the waveform, including vast amounts of information the ear cannot register. A lossy encoder analyzes the audio with a psychoacoustic model — an algorithm built from decades of hearing research — and identifies the components that are inaudible: sounds masked by louder nearby sounds, frequencies beyond hearing range, detail below the threshold of perception. It then re-encodes the audio with precision allocated according to audibility, spending its limited bit budget on what you will hear and starving what you will not. The result is a file a fifth to a tenth of the original size that sounds nearly identical. The key word is discarding: unlike lossless compression, the removed information is gone permanently — decoding a lossy file produces an approximation of the original PCM, not the original itself. That one-way nature drives every best practice around lossy audio.

The heart of every lossy codec is masking — the well-measured phenomenon that the ear fails to hear certain sounds in the presence of others. Frequency masking (simultaneous masking): a loud tone hides quieter sounds at nearby frequencies; a loud guitar note at 1 kHz makes soft content just above and below it imperceptible, so the encoder can represent that region crudely without audible harm. Temporal masking: a sudden loud event, like a snare hit, masks quieter sounds for a few milliseconds before it and up to 100-200 milliseconds after it, letting the encoder relax precision around transients. Absolute threshold of hearing: sensitivity varies across frequency — the ear is sharpest around 2-5 kHz and much less sensitive at the extremes — so very low and very high frequencies need less precision, and content above roughly 16-20 kHz can often be discarded outright. The encoder computes a masking threshold for each moment and frequency band: everything below the threshold is inaudible and can be quantized harshly or removed. That threshold map is effectively a live map of what your ears will actually notice.

Mechanically, most lossy codecs follow the same three-stage pipeline. First, transform: the audio is cut into short overlapping blocks and converted from the time domain to the frequency domain, typically with a Modified Discrete Cosine Transform (MDCT), turning each block into a set of frequency coefficients — how much energy exists at each pitch. Codecs switch between long blocks (better frequency resolution for steady tones) and short blocks (better time resolution for drum hits) to avoid smearing transients. Second, quantization: guided by the psychoacoustic model's masking threshold, each coefficient is rounded with variable precision — fine steps where the ear is sensitive, coarse steps where noise will be masked. This is where data is actually lost, and where the bitrate setting acts: more bits allow finer quantization everywhere. Third, packing: the quantized values are compressed further with lossless entropy coding (like Huffman coding) and wrapped into the file format with headers and metadata. Decoding reverses the packing and transform, but the quantization rounding is permanent — that is the 'loss' in lossy.

Four lossy codecs matter today, plus one legacy. MP3 (1993) is the original mainstream lossy format and still the most universally compatible file type in audio — every device made in twenty years plays it; it needs roughly 192-320 kbps for excellent quality. AAC is MP3's designed successor, delivering better quality at the same bitrate — a 128 kbps AAC roughly matches a 192 kbps MP3 — and is the standard for Apple Music, YouTube, and most streaming; hardware support is near-universal on modern devices. OGG Vorbis is the open-source, patent-free alternative that beats MP3 at equal bitrates; it found its home in games and open-source software but has weak Apple support. Opus (2012) is the current state of the art: the best quality per kilobit at nearly every rate, dominant in voice/video calling and much of YouTube's audio, limited mainly by patchy consumer-hardware support. WMA is Microsoft's largely obsolete entry, worth converting away from. The pattern across generations: each new codec's better psychoacoustic model buys the same quality at a lower bitrate.

Bitrate is the dial that controls how much data the encoder may keep, and therefore how aggressively it must prune. At high bitrates the encoder can represent everything above the masking threshold accurately, and the result is transparent — indistinguishable from the source for practically all listeners. As bitrate drops, the encoder is forced to quantize coarsely even in audible regions, and artifacts emerge. Rough MP3 reference points: 320 kbps and V0 are transparent for nearly everyone; 192 kbps is near-transparent for most listeners; 128 kbps is acceptable for casual listening but detectable on good equipment; 96 kbps and below shows clear degradation on music. Newer codecs shift the scale down: AAC is comfortable around 128-256 kbps, and Opus achieves near-transparency for music around 96-128 kbps and startling voice quality at 24-48 kbps. Variable bitrate (VBR) modes improve all of these by spending bits where content demands rather than uniformly. The right bitrate is the one where artifacts are inaudible in your real listening context — critical listening on monitors demands more than podcasts on a commute.

When a lossy encoder runs out of bits, the damage takes recognizable forms worth learning to hear. Pre-echo: a faint smear or hiss that arrives just before a sharp transient like a castanet click or snare hit, caused by quantization noise spreading across a long transform block — the classic MP3 artifact. Birdies or swirlies: warbly, watery, metallic noises on sustained tones and complex textures like applause, cymbals, or crowd noise, as marginal frequency components flicker in and out between blocks. Bandwidth limiting: a dull, muffled top end where the encoder simply discarded high frequencies to save bits — common at low bitrates. Stereo collapse: narrowed or unstable stereo image when joint-stereo tools are pushed hard. Applause, cymbals, harpsichord, and dense electronic textures are the torture tests where artifacts appear first, because their noise-like complexity gives the psychoacoustic model the least masking to hide behind. If a bitrate handles those cleanly on your equipment, ordinary material will be fine.

Lossy compression is safe once and dangerous repeatedly. Each encoding pass discards information according to its model's judgment of what is inaudible — but the judgments of successive passes do not line up. Re-encoding an MP3 to AAC, or even MP3 to MP3 at a different bitrate, decodes the already-approximated audio and throws away a fresh layer of detail, including parts the first pass had deliberately preserved. Artifacts that were individually below audibility stack until they are plainly audible: the second generation sounds slightly duller and swirlier, the fifth sounds obviously degraded, and a deliberately re-encoded file quickly turns to mush. This is generation loss, and it is why the cardinal rule of lossy audio is: encode from a lossless source, exactly once, as the final step. If you need the same audio in three lossy formats, make all three from the lossless master rather than chaining conversions. And when a lossy-to-lossy conversion is truly unavoidable, use a generously higher bitrate on the output to minimize the fresh damage.

'Transparent' is the term for a lossy encode that listeners cannot reliably distinguish from the lossless original, and it has a rigorous meaning grounded in blind testing. The standard method is an ABX test: you hear the original (A), the encoded version (B), and an unknown (X) that is one of the two, and must identify X repeatedly; statistically beating chance proves you can hear a difference, while failing to means the encode is transparent for you on that material and equipment. Decades of such tests underpin the received bitrate wisdom — for example that 320 kbps MP3 or 256 kbps AAC is transparent for nearly all listeners on nearly all content. Two honest caveats: transparency is personal and material-dependent (trained ears on revealing gear with torture-test tracks can detect what casual listening never will), and sighted comparisons are worthless because expectation bias overwhelms genuine audible differences. The practical takeaway: if a blind test cannot tell your chosen bitrate from lossless on your equipment, a higher bitrate buys you nothing but file size.

The two families of compression solve different problems. Lossy (MP3, AAC, Opus, Vorbis) optimizes for size: files shrink to a tenth of the original by discarding inaudible detail, at the cost of the loss being permanent and compounding across re-encodes. Lossless (FLAC, ALAC, WavPack) optimizes for fidelity: files shrink to roughly half by exploiting mathematical redundancy, and decoding returns the exact original PCM, bit for bit — the compression is fully reversible. The decision framework: audio you may ever edit, master, archive, or re-convert belongs in lossless, because it survives any number of future format changes; audio at its final destination — streaming to listeners, on a phone, in a podcast feed — belongs in lossy, because the size savings are enormous and a good encode is audibly indistinguishable in normal listening. The two work together, not in competition: keep the lossless master, distribute the lossy copy. What you should never do is treat a lossy file as a master, or 'upgrade' a lossy file to lossless expecting quality back — the discarded detail does not return.

Lossy is the right choice for essentially all delivery to human listeners. Streaming music and video: every major platform delivers lossy (AAC, Opus, Vorbis) because bandwidth matters and transparency is achievable. Podcasts: MP3 or AAC at modest bitrates is the standard, since voice compresses beautifully. Web audio and apps: small files mean fast loads. Phones and portable libraries: a lossy library fits ten times the music in the same storage. Email and messaging: attachments must be small. Voice recordings and meetings: intelligibility survives aggressive compression. The guiding principle: when the file has reached its final form and its job is to be heard — not edited, not archived, not re-encoded — lossy at an adequate bitrate is not a compromise but the correct engineering choice. The overwhelming majority of all audio consumed on earth is lossy, and at sensible bitrates nobody can tell.

Avoid lossy anywhere the audio has a future beyond playback. Archival: once data is discarded it is gone forever, so masters and irreplaceable recordings belong in lossless (FLAC) or uncompressed (WAV) — storage is cheap, re-recording is impossible. Production: never edit, mix, or process lossy files if a lossless source exists; every render-and-re-encode cycle adds a generation of loss, so stems, session files, and masters stay lossless until the final export. Transcoding chains: converting one lossy format to another compounds artifacts — go back to the lossless original instead. Critical listening libraries: if you have revealing equipment and care, lossless costs only disk space. Sample libraries and sound design assets: they will be pitched, stretched, and processed, which exposes and amplifies lossy artifacts. The rule of thumb: lossy is a one-way exit door for finished audio — walk through it last, and keep a lossless copy of everything on the other side.

The golden path is lossless-to-lossy: encode your WAV or FLAC master to MP3, AAC, or Opus at the delivery stage, choosing 192-320 kbps for music or 96-128 kbps for voice, and keep the lossless original. Converting lossy-to-lossless (MP3 to FLAC) is harmless but pointless — it produces a big file with the MP3's losses baked in, not restored quality. Converting lossy-to-lossy (WMA to MP3, OGG to AAC) is sometimes a practical necessity for device compatibility; accept the small generational loss and use a healthy output bitrate to limit it. Converting lossy-to-uncompressed (MP3 to WAV) is the right move before editing, since it prevents further loss during processing even though it cannot recover what is gone. AudioUtils performs all of these conversions entirely in your browser using FFmpeg WebAssembly — no upload, no signup, no software install, and your files never leave your device.