Free Lossless Audio Codec

How FLAC compression works

FLAC uses a prediction-plus-residual architecture. For each block of audio samples, the encoder fits a predictor — either a simple fixed polynomial or a linear predictive coding (LPC) model — to the waveform. The predictor estimates each sample based on preceding samples. The encoder then stores only the residual: the difference between the predicted value and the actual value. Because a good predictor tracks the signal closely, residuals cluster near zero and compress efficiently.

Those residuals are encoded using Rice coding (a variant of Golomb coding optimized for geometric distributions). The encoder tests multiple Rice parameters per partition and selects the one that produces the smallest output. This is where FLAC spends most of its computation during encoding.

Each channel within a frame can use a different coding strategy. FLAC supports independent coding (each channel compressed separately), left-side, right-side, and mid-side stereo decorrelation. Mid-side mode stores the sum and difference of left and right channels; when the channels are similar — common in most stereo recordings — the difference signal is small and compresses tightly.

Frame structure

A FLAC stream is a sequence of frames, each self-contained. Every frame begins with a sync code (0xFFF8 or 0xFFF9), followed by a header that records the block size, sample rate, channel assignment, bit depth, and a frame or sample number. After the header, each channel gets its own subframe containing the predictor order, predictor coefficients (for LPC), and the Rice-coded residuals. A CRC-8 protects the frame header; a CRC-16 covers the entire frame. If a single bit flips in storage or transmission, the decoder detects it at the frame level without corrupting the rest of the stream.

Block size is configurable. The default is 4096 samples for most sample rates. Larger blocks give the predictor more context and marginally better compression. Smaller blocks reduce decode latency and improve seek granularity.

Compression levels and diminishing returns

FLAC offers compression levels 0 through 8. Level 0 uses fixed predictors, small block sizes, and minimal partition search. Level 8 uses high-order LPC predictors (up to order 12), exhaustive partition search, and larger blocks. FLAC compression level 0 vs level 8 differs by only 3–5% in file size but encoding time increases 8–10x — level 5 (the default) hits the knee of the curve. Beyond level 5, each increment buys fractions of a percent in size reduction while doubling or tripling encode time. For archival encoding where time is irrelevant, level 8 saves a few gigabytes across a large library. For real-time recording or batch processing, level 5 is the practical ceiling.

Seek table and random access

FLAC files include an optional seek table in the metadata — a list of sample numbers mapped to byte offsets within the stream. Players use the seek table to jump to any point in the file without decoding from the start. Without a seek table, seeking still works (the decoder scans for sync codes), but it is slower on large files or network streams. Most encoders generate a seek table by default with one entry every 10 seconds.

Metadata: Vorbis comments and album art

FLAC uses the Vorbis comment system for metadata — the same key-value tag format used by Ogg Vorbis. Tags are UTF-8 strings, case-insensitive, and support arbitrary fields beyond the standard ARTIST, ALBUM, TITLE set. ReplayGain values are stored as Vorbis comments.

Album art is embedded via METADATA_BLOCK_PICTURE, a binary metadata block that stores the image type (front cover, back cover, etc.), MIME type, description, dimensions, and the raw image data. Multiple picture blocks are allowed. Some older software writes album art as a Vorbis comment with base64-encoded image data instead — this works but inflates the metadata size by roughly 33%.

Subset mode and streaming

FLAC defines a "subset" — a restricted set of encoding parameters that all conformant decoders must support. Subset mode limits block size to 4608 samples at 48 kHz and below (16384 above 48 kHz), predictor order to 12, and Rice partition order to 8. Files encoded within subset constraints are guaranteed to stream and decode on any FLAC-compatible device. Non-subset files may use higher predictor orders or unconventional block sizes that some hardware decoders reject.

Ogg FLAC

FLAC can be encapsulated in an Ogg container, producing Ogg FLAC files (typically .oga extension). Ogg FLAC adds Ogg's framing, multiplexing, and chained stream features. This matters for internet radio and streaming servers built on the Icecast/Ogg infrastructure. Native FLAC is more common for file storage; Ogg FLAC fills a niche in live streaming.

FLAC vs alternatives

ALAC (Apple Lossless) compresses to similar ratios — within 1–2% of FLAC on most material. ALAC's advantage is native support in Apple's ecosystem: iTunes, Music app, iPhone, AirPlay. FLAC is supported everywhere else and has broader hardware player compatibility (most dedicated audio players, Android, car systems, Sonos, and all major streaming DACs).

WavPack offers a hybrid mode that produces a lossy file plus a correction file; combining both recovers the lossless original. This is unique among lossless codecs. WavPack also supports 32-bit float samples natively. Compression ratios are comparable to FLAC, sometimes 1–2% better on specific material.

APE (Monkey's Audio) compresses 2–5% smaller than FLAC on average but decodes significantly slower. High-compression APE files can strain low-power hardware decoders. The format also lacks a seek table, making random access expensive on large files.

Limitations

FLAC has no DRM mechanism. This is by design — the format is open and patent-free — but it means services requiring copy protection cannot use it for distribution.

Channel count maxes out at 8. For immersive audio formats like Dolby Atmos or 22.2 surround, FLAC cannot carry the full bed and object layout in a single stream.

There is no lossy mode. If you want lossy compression from the same ecosystem, Opus (for general audio) or Vorbis (legacy) are the Xiph.org alternatives. FLAC is strictly lossless.

Maximum bit depth is 32 bits per sample. Maximum sample rate is 655,350 Hz. In practice, 24-bit/192 kHz is the highest common configuration, and FLAC handles it without issue.

When to choose FLAC

Use FLAC for music archival, library storage, and any workflow where preserving the original recording matters. It is the standard lossless format for Linux, Android, and most non-Apple hardware.

Use WAV when you need uncompressed PCM for DAW editing, plugin compatibility, or broadcast ingest chains that expect raw audio. WAV avoids any decode overhead but costs 35–50% more storage.

Use ALAC if your playback chain is exclusively Apple hardware and software. The compression and quality are equivalent to FLAC; only the compatibility matrix differs.

Use Opus at 128–160 kbps for transparent lossy quality at a fraction of the file size. Opus outperforms AAC and MP3 at every bitrate and is the correct choice when storage or bandwidth constraints make lossless impractical.