Are all audio codecs lossy?

No. Lossless codecs such as FLAC recover the original supplied digital samples exactly. Other codecs deliberately trade exactness for smaller representations.

Is WAV always an audio codec?

WAV is a container format. It commonly contains PCM, but the container name and the audio representation are different layers.

Will changing the extension convert a file?

No. Renaming does not decode and re-encode the audio. A real conversion tool must interpret the existing format and write the chosen new one.

AI Glossary / MLOps and deployment / Beginner

Audio codec

Definition

An audio codec is a method, together with an implementation of that method, for encoding and decoding digital audio. It defines how sound samples are represented for storage or transmission and how a decoder reconstructs playable audio. Many codecs compress data, either losslessly or lossily.

What it means

Codec combines the ideas of coding and decoding. An encoder creates the coded representation; a decoder interprets it. A codec can be implemented in software, hardware or both.

A lossless codec preserves the original digital samples exactly when decoded correctly. A lossy codec produces an approximation, usually to reduce the amount of data further. Neither type recreates information absent from its input samples.

A codec is different from a container. The codec describes the audio coding, while a container packages coded audio with timing, metadata and possibly other streams. An extension alone may not identify the required parameters.

How it works

A recording system first obtains digital samples, characterized by properties such as sample rate and channel count. The encoder processes those samples into a coded stream using the selected method and settings.

A lossless approach can predict samples and efficiently code the exact prediction errors. FLAC is a documented example of lossless audio coding. A lossy approach may represent selected signal features or quantize values according to its design. Opus is an example designed for interactive speech and audio.

The decoder reads the encoded information and produces samples. Playback may then involve resampling, mixing and conversion to an analog signal. These additional steps should not be confused with the codec itself.

A working pipeline must agree on format, supported modes and framing. For communication, packetization, buffering and missing-packet handling also affect the listener's experience. Codec quality is only one part of end-to-end sound quality.

Three practical examples

1. Voice calls: A communication application encodes microphone audio for transmission and decodes received packets for playback. Low delay can matter as much as file size.

2. Audio archives: A team may choose lossless coding when it needs exact recovery of the supplied samples for later editing or analysis. Metadata and backups still need separate attention.

3. Speech AI: A transcription service receives audio in a supported coded format, decodes it, and prepares samples for its model. Incorrect sampling information can damage results even when the speech model is otherwise suitable.

Think of it this way

Think of a codec as an agreed way to pack and unpack a recording. Lossless packing returns every original item; lossy packing keeps an approximation selected according to a rule. The container is the labeled box holding the packed material.

Different codecs and settings can remove different details. Assess sound through listening and task-specific tests, not the label alone.

Worked example

Consider one minute of stereo linear PCM audio sampled at 48,000 samples per second, with 16 bits per sample in each channel. The raw payload rate is 48,000 × 16 × 2 = 1,536,000 bits per second.

Over 60 seconds, that is 92,160,000 bits, or 11,520,000 bytes after dividing by eight. Using decimal units, the sample payload occupies 11.52 MB. This excludes any file header or metadata.

Suppose a selected lossy encoding averages 128,000 bits per second for the complete stereo stream. Its one-minute coded payload is approximately 128,000 × 60 / 8 = 960,000 bytes, or 0.96 MB. The payload ratio is 11.52/0.96 = 12 to 1.

That calculation says nothing by itself about audible quality or transcription accuracy. It also assumes the stated average rate; variable-rate output can differ from a nominal setting. Container and network overhead add bytes.

A lossless file's size cannot be inferred from this lossy target. Its compression depends on the audio's structure, but successful decoding must reproduce the original supplied samples exactly.

Illustrative case study

Illustrative case study - fictional

A fictional training company records interviews and publishes edited lessons. It keeps a lossless editing copy, creates a smaller delivery version, and sends an approved version to its transcription workflow.

When transcripts show errors, the team checks the actual decoded waveform rather than assuming that a familiar extension guarantees correct audio. It finds a channel-handling mistake that mixed an interviewer channel incorrectly. Fixing the preparation step is evaluated separately from changing codec settings.

The company documents each version's purpose. This invented scenario reports no measured improvement.

Common misconceptions and mistakes

Codec, container and filename extension are related but different concepts.
A higher bitrate is not a universal quality ranking across different codecs and settings.
Converting lossy audio into a lossless format preserves the decoded approximation, not the original discarded detail.
Resampling to a higher sample rate does not automatically create new captured frequencies.
Repeated lossy transcoding can change the signal further even when file size stays similar.

Business use

Audio-codec choices affect storage, network use, latency, device compatibility and downstream analysis. Begin with the intended purpose: live conversation, listening delivery, editing or exact sample preservation can have different priorities.

Test representative speech, music and noise rather than relying on a single clean sample. For AI services, assess recognition or classification using the same decoding and preparation path used in production. Document sample rate, channels, codec settings and conversions so results can be reproduced. Retaining an appropriate source recording avoids repeatedly rebuilding new versions from an already degraded delivery copy.

Check your understanding

A lossy recording is decoded and saved using a lossless codec. Has the original pre-compression recording been recovered?

Answer: No. The lossless file can preserve the decoded samples exactly, but it cannot restore information discarded by the earlier lossy encoding.

Related terms

Lossless compression
audio container
pulse-code modulation
bitrate
sample rate
bit depth
transcoding
speech recognition

Learn next

Inspect a recording's container, codec, sample rate and channel count separately. Calculate its uncompressed PCM payload, then compare actual encoded sizes while keeping duration and channel configuration constant.

Sources

Sources: IETF RFC 9639, Free Lossless Audio Codec: https://www.rfc-editor.org/rfc/rfc9639 ; Opus project technical overview: https://www.opus-codec.org/ . The FLAC introduction and lossless/lossy definitions and the Opus overview were reviewed. The size calculation is an original hypothetical example, not a codec benchmark.

Previous: Auction algorithm Next: Audio inpainting