Lossless vs Lossy Audio: The Complete Guide

Every audio format on a hard drive, every stream from Apple Music or Spotify, every voice memo on a phone — all of it falls on one side of a single dividing line. Lossless compression keeps every audio sample exactly as recorded, just packed into fewer bytes. Lossy compression discards information the encoder predicts you cannot hear, achieving file sizes that lossless cannot match. Understanding which side a format falls on, and what the math behind each side actually does, is the foundation of every other audio decision: whether to archive in FLAC or WAV, whether 256 kbps AAC is enough for distribution, why transcoding lossy files is something you avoid.

This guide goes deeper than the surface treatment. It covers the entropy-coding math that makes lossless work, the psychoacoustic masking that lossy codecs exploit, the bitrates at which trained listeners can ABX-distinguish lossy from lossless, the specific codecs in each category, and the workflow rules that prevent storage waste and quality loss.

What Lossless Compression Actually Does

Lossless compression treats audio the way ZIP treats a text file. The bytes in are different from the bytes out — but when you decompress, you recover the exact original byte sequence. Every audio sample, every bit of dynamic range, every frequency captured at recording, returns intact.

Three classes of algorithms make this possible.

Predictive coding. Audio samples are correlated — sample N+1 is usually close to sample N. A linear predictor models the next sample as a weighted combination of the previous few samples, then stores only the residual (the difference between prediction and actual). For typical music, residuals are much smaller than the original samples, so they need fewer bits to encode. FLAC uses fixed and Linear Predictive Coding (LPC) predictors of order 1 through 32; the encoder chooses per-block which order minimizes residual size.

Entropy coding. The residuals from prediction are then compressed with a variable-length code that gives shorter codes to more common values. FLAC uses Rice coding, a special case of Golomb codes optimized for residuals with a Laplacian distribution. ALAC uses a similar approach. Generic Huffman coding is used elsewhere — for example, in the entropy stage of lossless image codecs like PNG.

Channel decorrelation. Stereo audio has two correlated channels. Encoding the sum (L+R)/2 and the difference (L-R) instead of L and R separately exploits this correlation, since most signals are dominated by the sum. FLAC supports four channel modes: independent stereo, left+side, right+side, and mid+side, choosing per block.

The result for typical music: FLAC files are 50-65 percent of the equivalent uncompressed WAV size. A 4-minute CD-quality track is about 42 MB as WAV and 23-28 MB as FLAC, with bit-perfect equivalence. The compression ratio depends on the source — sparse acoustic music compresses better than dense electronic mixes. Total silence compresses extremely efficiently because the predictor's residuals are zero.

What Lossy Compression Actually Does

Lossy codecs use a fundamentally different approach. They model how human hearing perceives sound and discard information the model judges to be inaudible. The discarded data cannot be recovered. The decoded output is similar to the source — sometimes perceptually indistinguishable, sometimes obviously degraded — but never bit-identical.

The core tool is the psychoacoustic model. Two phenomena drive it.

Frequency masking. A loud tone at one frequency raises the audibility threshold for nearby frequencies. A 1 kHz sine wave at 80 dB hides any tone within 50-150 Hz of it that is more than about 30 dB quieter. The encoder analyzes each frame, identifies the loud "masker" tones, computes the masking thresholds in surrounding bands, and quantizes the masked bands more harshly — using fewer bits because the noise introduced by quantization is hidden under the masker.

Temporal masking. A loud transient masks quieter sounds for roughly 5 ms before (pre-masking) and 100-200 ms after (post-masking). The encoder allocates fewer bits to those temporal windows. This is why drum hits compress efficiently: everything in the few hundred milliseconds around them is masked.

The encoder transforms the time-domain signal into the frequency domain (typically via Modified Discrete Cosine Transform), applies the masking model to determine bit allocation per frequency band, quantizes each band, then uses entropy coding to compress the quantization indices. The output is a stream of frames containing quantized frequency-domain coefficients plus side information.

Lossy codecs achieve compression ratios that lossless cannot approach. A 4-minute song at 128 kbps MP3 is about 3.84 MB — 9 percent of the uncompressed size, less than a third of typical FLAC. At 64 kbps Opus the same song is 1.92 MB, 4.5 percent of the original.

The Numbers Compared

Practical compression ratios for typical music:

• WAV / AIFF (uncompressed PCM). 100 percent. CD quality is 1,411 kbps stereo.
• FLAC / ALAC (lossless compression). 50-65 percent. Roughly 700-1,000 kbps for CD-quality material.
• 320 kbps MP3 (high-bitrate lossy). About 23 percent. Perceptually transparent for nearly all listeners.
• 256 kbps AAC (Apple Music default). About 18 percent. Perceptually transparent for nearly all listeners.
• 128 kbps AAC. About 9 percent. Approaching transparency on typical material.
• 96 kbps Opus. About 7 percent. Approaching transparency, the current state of the art.
• 64 kbps Opus. About 4.5 percent. Audible but acceptable for speech and casual music.

The full audio bitrate explainer covers the numbers in context. The summary: lossless compresses by half; lossy compresses by 10x to 25x.

When Listeners Can Actually Tell the Difference

The practical question is not whether lossy codecs lose data — they always do — but whether listeners can hear it. Public ABX testing (the gold standard is Hydrogen Audio's listening tests) gives rough thresholds.

For trained listeners on revealing playback equipment in quiet rooms:

• MP3 at 320 kbps. Distinguishable from FLAC on roughly 5 percent of trials. Statistically detectable but not practically audible.
• MP3 at 192 kbps. Distinguishable on roughly 20-40 percent of trials, depending on material. Cymbals, reverb tails, and complex polyphony reveal it.
• MP3 at 128 kbps. Audibly different on most material. Pre-echo on transients, swirly artifacts on cymbals.
• AAC at 128 kbps. Approaches transparency. Better than MP3 at the same bitrate.
• Opus at 96 kbps. Approaches transparency. The current state of the art for stereo music.

For typical listeners on consumer playback equipment in normal environments:

• 128 kbps AAC sounds the same as the source on Bluetooth earbuds in a coffee shop.
• 192 kbps MP3 is the practical floor for music on car stereos and home speakers.
• 64 kbps speech codecs are fine for phone calls and dictation.

The equipment is often the limiting factor. Laptop speakers cannot reproduce the high-frequency detail that distinguishes 128 kbps from FLAC. Good headphones in a quiet room expose every artifact down to bit-perfect lossless.

The Format Lineup

The major codecs sorted by category:

Lossless audio formats:

• FLAC — Free Lossless Audio Codec. Open source, royalty-free, broad device support. The standard choice for archival and high-quality streaming.
• ALAC — Apple Lossless Audio Codec. Functionally similar to FLAC, packaged in MPEG-4 (M4A) containers. The Apple ecosystem's lossless format. Compare in the FLAC vs ALAC explainer.
• WAV — Microsoft's uncompressed PCM container. Bit-perfect, no compression at all. The studio standard for recording and editing.
• AIFF — Apple's uncompressed PCM container. Functionally identical to WAV with different byte order.
• WavPack — Hybrid lossless/lossy codec. Niche in 2026.
• APE (Monkey's Audio) — Lossless codec popular in some communities. Less universal device support than FLAC.

Lossy audio formats:

• MP3 — MPEG-1 Audio Layer III. Universal hardware compatibility. Less efficient than modern codecs.
• AAC — Advanced Audio Coding. The successor to MP3, used by Apple Music, YouTube, and most streaming services. Roughly 30 percent more efficient than MP3.
• Opus — The IETF's open-source codec. The current efficiency leader. Used by WhatsApp, Discord, YouTube, and modern web standards.
• Vorbis — Older open-source codec. Used by Spotify free tier.
• WMA — Windows Media Audio. Effectively obsolete in 2026.
• HE-AAC, xHE-AAC — AAC profiles optimized for low-bitrate streaming and broadcast.

PCM data inside WAV or AIFF is technically uncompressed, not "lossless compressed" — but it sits in the lossless camp because no audio data is discarded. See the PCM audio explainer for the underlying digital audio model.

The One-Way Street

The single most important rule in audio file management: lossy compression is permanent. You can always encode lossless to lossy. You can never recover lossless from lossy.

When you convert WAV to MP3, the encoder analyzes the PCM source and produces a smaller file with discarded information that the psychoacoustic model judged inaudible. When you then convert that MP3 back to WAV, the decoder reconstructs PCM samples from the MP3 stream — but those samples reflect what the MP3 actually encoded, not the original source. The WAV will be the same size as the original WAV but will contain the lossy artifacts of the MP3. The lost data stays lost.

The same applies to lossless-to-lossless conversions, but in reverse: those are bit-perfect round trips. Convert WAV to FLAC and back to WAV, and you get the original WAV byte-for-byte. The intermediate FLAC was just a smaller package.

Generational Loss in Lossy-to-Lossy Transcoding

Each lossy encode applies a fresh psychoacoustic pass. Quantization noise compounds. Even high-bitrate transcodes — 320 kbps MP3 to 256 kbps AAC, for instance — measurably degrade the signal compared to encoding the AAC directly from a lossless source.

The cumulative damage is sometimes audible after two transcodes and almost always audible after three. This is why podcast workflows that re-encode multiple times produce noticeably worse audio than single-pass workflows from the source. Always start from the highest-quality source available — ideally lossless.

The Workflow That Eliminates Most Regret

A simple rule covers nearly every consumer and prosumer audio scenario:

1. Record in lossless. WAV or AIFF for studio work, FLAC if storage is tight. 2. Edit in lossless. DAWs handle WAV most efficiently; some prefer FLAC or AIFF. 3. Archive in lossless. FLAC saves about half the space of WAV with zero quality loss. The FLAC vs WAV for production comparison covers when each is preferable. 4. Distribute in lossy. MP3 for universal compatibility; AAC for better quality on modern devices; Opus for the open web. 5. Never delete the lossless masters. They are your only option for re-encoding to future formats.

The workflow scales from podcast production to recording studios. The cost of keeping FLAC or WAV masters is storage, which is cheap. The cost of throwing them away is irreversible, and modern lossy codecs improve every few years — material you encoded as 192 kbps MP3 in 2010 could be transparent at 96 kbps Opus today, but only if you still have the lossless source.

Specific Codec Recommendations

The right choice depends on the use case.

• Studio recording and mixing. WAV at 24-bit / 48 kHz or 24-bit / 96 kHz. The DAW prefers it, and bit depth above 16 gives mixing headroom.
• Music archive (personal library). FLAC at the source resolution. Compresses to roughly half the WAV size. The M4A vs FLAC quality comparison addresses the frequent question about Apple ecosystems specifically.
• Music streaming distribution. AAC at 256 kbps for Apple-friendly delivery, MP3 at 320 kbps for maximum compatibility, Opus at 96-128 kbps for web-native applications.
• Podcast distribution. MP3 at 96-128 kbps mono for speech, 128-192 kbps stereo if music beds matter. Apple Podcasts supports both MP3 and M4A AAC.
• Voice messages and dictation. Opus at 32-64 kbps, or if Opus is unavailable, AAC at 64 kbps mono.
• Audio for video. AAC at 128-192 kbps for delivery; the video codec dominates the bitrate budget.

Bottom Line

Lossless compression preserves every bit; lossy compression discards data the encoder model judges inaudible. FLAC and ALAC compress to roughly half the WAV size with bit-perfect reconstruction. MP3, AAC, and Opus compress to roughly 5-25 percent of WAV size with audible artifacts at low bitrates and effective transparency at moderate bitrates (192 kbps MP3, 128 kbps AAC, 96 kbps Opus). The cardinal rule is the one-way street: archive lossless, distribute lossy, never delete the lossless originals. Following that rule eliminates nearly every audio quality regret a user can have, scales from podcast production to recording studios, and costs only storage — which is the cheapest part of any modern audio workflow.