The problem: CD-quality audio is enormous
A 3-minute song at CD quality is about 30 megabytes. That is 10 megabytes per minute, or 170 kilobytes per second. Streaming a song at that rate requires a 1.4 Mbps connection with zero overhead — feasible on wifi, painful on mobile data, impossible on a weak signal. The reason is simple: CD-quality audio is uncompressed PCM (pulse-code modulation), and the math of uncompressed audio multiplies fast.
Audio file size = sample rate × bit depth × channels × duration. For CD quality: 44,100 samples per second × 16 bits per sample × 2 channels = 1,411,200 bits per second, or 176,400 bytes per second. Over 180 seconds (3 minutes), that is 31,752,000 bytes — about 30 MB. Every second of CD audio is 176 KB. Every minute is 10 MB.
The Audio Compressor reduces file size by lowering these four parameters: sample rate, bit depth, channel count, and (optionally) applying peak normalization. It outputs WAV PCM, which is uncompressed but smaller because it contains fewer samples. The reduction is mathematical — fewer bits per second — not codec-based, so the result is predictable and lossless at the chosen parameters (no codec artifacts, just lower fidelity).
Fastest path
Open the Audio Compressor, drag your audio file (MP3, WAV, OGG, M4A, AAC, FLAC) into the upload zone. Pick a sample rate (44100 for CD, 22050 for FM, 16000 for voice, 8000 for telephone). Pick bit depth (16-bit for standard quality, 8-bit for 50 percent smaller files at lower quality). Toggle mono if you do not need stereo. Toggle normalize if you want the loudest peak at maximum level. The tool shows the estimated output size and the compression ratio before you process. Click compress, then download the WAV.
Sample rate: the Nyquist limit
Sample rate is how many times per second the audio waveform is measured. CD audio uses 44,100 Hz — 44,100 measurements per second. Each measurement is a single number representing the instantaneous amplitude of the sound wave.
The Nyquist-Shannon sampling theorem says the sample rate must be at least twice the highest frequency you want to capture. Human hearing tops out at about 20,000 Hz, so 44,100 Hz (just above 2 × 20,000) captures the full audible range with a small margin. This is why CD audio was standardized at 44.1 kHz in the 1980s — it could store the full audible spectrum on the equipment of the time.
Lower the sample rate and you lose the high frequencies:
| Sample rate | Highest frequency captured | Use case |
|---|---|---|
| 44,100 Hz | 22,050 Hz | CD quality, music |
| 22,050 Hz | 11,025 Hz | FM radio quality, background music |
| 16,000 Hz | 8,000 Hz | Voice (speech is mostly 100-8,000 Hz) |
| 8,000 Hz | 4,000 Hz | Telephone quality (muffled but intelligible) |
Telephone audio at 8,000 Hz captures frequencies up to 4,000 Hz. Human speech is mostly between 300 and 3,400 Hz, which is why telephone audio is intelligible — the range that matters for understanding speech is preserved. What you lose is the high-frequency sibilance (the "s" and "t" sounds above 4,000 Hz) and the fullness of music. A song at 8,000 Hz sounds muffled, like it is playing through a wall.
Cutting sample rate halves file size at each step. A 3-minute stereo 16-bit file is 30 MB at 44,100 Hz, 15 MB at 22,050 Hz, 10.9 MB at 16,000 Hz, and 5.5 MB at 8,000 Hz. The reduction is linear: half the samples, half the bytes.
Bit depth: dynamic range
Bit depth is how many bits are used to represent each sample. CD audio uses 16-bit samples — each measurement is a number between -32,768 and 32,767. Telephone audio historically used 8-bit samples (with a non-linear companding scheme that effectively gave 12-13 bits of dynamic range, but the storage was 8 bits per sample).
Bit depth controls dynamic range — the gap between the quietest and loudest sound the format can represent. The rule is 6 dB of dynamic range per bit:
| Bit depth | Dynamic range | Use |
|---|---|---|
| 24-bit | 144 dB | Studio recording, mastering |
| 16-bit | 96 dB | CD quality, consumer playback |
| 8-bit | 48 dB | Telephone, low-quality voice |
A 16-bit recording can represent sounds from a whisper (about 0 dB) to a jackhammer at 1 meter (about 100 dB) without clipping or noise. 8-bit has only 48 dB of range — the quietest sounds disappear into quantization noise (a hiss), and the loudest sounds clip earlier. 8-bit is fine for voice where the dynamic range is small and consistent; it is bad for music where quiet passages need to stay quiet and loud passages need to stay loud.
Cutting bit depth halves file size. A 3-minute stereo 44,100 Hz file is 30 MB at 16-bit and 15 MB at 8-bit. The cost is 48 dB of dynamic range — the quietest sounds hiss, the loudest sounds distort sooner.
Channels: mono vs stereo
Stereo audio has two channels — left and right — which doubles the file size. Mono has one channel. Converting stereo to mono halves the file size and loses the stereo image (the spatial placement of sounds left and right).
For music, stereo matters. A recording of a band has the guitar on the left, the drums in the center, the keyboard on the right — the spatial information is part of the experience. Collapsing to mono sums the channels and flattens the image. The music is still recognizable, but it sounds flat, like it is coming from a single point.
For voice, mono is fine. A single speaker has no stereo image to lose. Podcasts, audiobooks, voice memos, and most spoken-word content are mono or should be — there is no benefit to stereo and a 50 percent file-size cost. A 3-minute voice memo at 16,000 Hz, 16-bit, mono is 5.5 MB. The same in stereo is 11 MB, with no perceptible quality gain.
The full reduction: 95 percent
Combining all three reductions — sample rate, bit depth, and channels — multiplies the savings:
| Configuration | Size (3 min) | Reduction |
|---|---|---|
| 44,100 Hz, 16-bit, stereo (CD) | 30 MB | 0% (baseline) |
| 22,050 Hz, 16-bit, stereo | 15 MB | 50% |
| 22,050 Hz, 16-bit, mono | 7.5 MB | 75% |
| 16,000 Hz, 16-bit, mono | 5.5 MB | 82% |
| 8,000 Hz, 16-bit, mono | 2.7 MB | 91% |
| 8,000 Hz, 8-bit, mono | 1.4 MB | 95% |
A 95 percent reduction takes a 30 MB CD-quality file to 1.4 MB. The trade-off is telephone quality — intelligible speech but muffled music, audible hiss on quiet passages, no stereo. For a voice memo or a podcast clip that needs to fit in an email attachment, that is the right trade-off. For a song you want to enjoy, it is not.
Why the output is WAV, not MP3
The tool outputs WAV PCM rather than MP3 or AAC. This is a deliberate choice with a specific reason: encoding MP3 in a browser requires a library (like lamejs or a WASM build of LAME), and the licensing history of MP3 made browser-native encoding uncommon until recently. WAV encoding is straightforward — a 44-byte header followed by raw sample data — and can be implemented in 50 lines of JavaScript.
The trade-off is that WAV is uncompressed. A 1.4 MB 8 kHz mono 8-bit WAV file contains the same audio as an MP3 at 32 kbps, which would be about 720 KB — half the size. Codec-based compression (MP3, AAC, Opus) achieves much higher ratios than PCM parameter reduction because it discards information the ear cannot hear (psychoacoustic modeling) rather than uniformly reducing the sample grid.
If you need MP3 output, the tool's FAQ suggests downloading the WAV and using FFmpeg externally: ffmpeg -i input.wav -b:a 64k output.mp3. A 64 kbps MP3 of voice audio is about the same quality as the 8 kHz mono 16-bit WAV but roughly half the file size. For music, 128 kbps MP3 is near-transparent to most listeners at 1/11 the size of CD-quality PCM.
Peak normalization
The tool's normalize option scales the audio so the loudest peak reaches 0 dBFS (decibels relative to full scale) — the maximum level the format can represent without clipping. This maximizes perceived loudness without distortion.
Peak normalization is not the same as loudness normalization (which targets a perceived loudness like -14 LUFS, used by Spotify and YouTube). Peak normalization guarantees no clipping but does not match the loudness of other tracks. A quiet recording normalized to 0 dBFS peaks at full scale but may still sound quieter than a heavily compressed modern recording that also peaks at 0 dBFS.
For the tool's use case — compressing voice or music for size, not for loudness matching — peak normalization is the right choice. It ensures the output uses the full dynamic range of the format, so you do not waste bits on a recording that peaks at -12 dBFS (using only a quarter of the available amplitude range).
Gotchas
- The tool outputs WAV, not MP3. WAV is uncompressed, so the size reduction comes from fewer samples (lower sample rate, fewer bits, fewer channels), not from codec compression. A 64 kbps MP3 of the same audio is roughly half the size of an 8 kHz mono 16-bit WAV. If you need MP3, download the WAV and convert with FFmpeg. The tool's FAQ acknowledges this.
- Resampling is lossy even at high sample rates. When the tool resamples a 48,000 Hz file to 44,100 Hz, it uses OfflineAudioContext's resampler, which applies an anti-aliasing filter that can introduce minor artifacts. For most audio this is inaudible, but for critical listening, resample once at the target rate and avoid repeated resampling.
- 8-bit audio has audible quantization noise. The 48 dB dynamic range of 8-bit means quiet passages have a hiss. This is acceptable for voice (the speech is louder than the noise floor) but noticeable on music. If the source has quiet passages (a classical recording with a hushed section), the hiss will be audible during those passages.
- Peak normalization does not match loudness across tracks. Two files both normalized to 0 dBFS can sound different in loudness because of dynamic range compression in the source. If you are producing a playlist, use a loudness-normalization tool (targeting -14 LUFS) rather than peak normalization.
- Mono downmix can cause phase cancellation. If the left and right channels are out of phase (a stereo effect, or a miswired cable), summing them to mono cancels the out-of-phase components and can dramatically reduce volume or eliminate certain sounds. The tool does not check for phase alignment. If a mono conversion sounds wrong, the source may have phase issues — check the original in mono before blaming the tool.
Summary
- Audio file size is sample rate × bit depth × channels × duration. CD-quality 3-minute stereo audio is 30 MB. Reducing any of the four parameters reduces the file proportionally. Combining all reductions — 8,000 Hz, 8-bit, mono — takes a 30 MB file to 1.4 MB, a 95 percent reduction.
- Sample rate must be at least twice the highest frequency to capture (Nyquist theorem). 44,100 Hz captures the full audible range; 8,000 Hz captures up to 4 kHz, which is why telephone audio sounds muffled. Bit depth gives 6 dB of dynamic range per bit: 16-bit has 96 dB, 8-bit has 48 dB (audible hiss on quiet passages).
- Stereo doubles file size and matters for music (spatial image) but not for voice. Mono is correct for podcasts, voice memos, and spoken-word content. The tool outputs WAV PCM — uncompressed but smaller because it has fewer samples. MP3 would be smaller still (codec compression), but browser-based MP3 encoding requires a library the tool does not include.
- Peak normalization scales the loudest peak to 0 dBFS, maximizing perceived loudness without clipping. It is not loudness normalization (which targets a perceived loudness like -14 LUFS). For a playlist, use loudness normalization; for a single file, peak normalization is sufficient.
- Use the Audio Compressor for PCM parameter reduction, the Audio Converter for format changes, the Audio Merger for combining files, and the Video to GIF Converter for video-size reduction.