Reading the numbers

LoudSense measures per ITU-R BS.1770 and EBU R128, the same standards a broadcast loudness meter uses. Here is what each number is and what to do with it.

Peak (dBFS)

The largest single sample in the file, relative to full scale. 0 dBFS is the biggest value a fixed-point file (16 or 24 bit) can hold. A peak above 0 dBFS is possible only in float data — a 32-bit float WAV, or a decoded MP3 or AAC. Those samples really are over full scale, and any fixed-point stage will clip them: a 16/24-bit export, the DAC, the phone's decoder.

True peak (dBTP)

The waveform between the samples can be higher than any sample. LoudSense reconstructs it by oversampling and reports the highest point — never lower than the sample peak. A sine at a quarter of the sample rate, offset by 45°, has samples at −3 dBFS and a true peak at 0 dBTP — the classic case. Platforms and encoders react to the true peak, which is why every target has a dBTP ceiling.

Clips

A run of consecutive samples at full scale counts as one clip; the positions are drawn as red lines. Twenty clipped samples in a row are one clip, not twenty. A handful in a drum hit is often inaudible; hundreds across a podcast mean the limiter or the encoder ran out of headroom.

LUFS-M, LUFS-S, LUFS-I

Loudness in LUFS (loudness units relative to full scale) uses the K-weighting filter — a high-pass at 38 Hz and a high shelf above 1.7 kHz — so the number tracks how loud something sounds, not how big the waveform is.

• LUFS-M (momentary) — a 400 ms window. The orange curve on the panel; the stats row shows its maximum, found at 10 ms resolution so a short burst between two curve points is not missed.
• LUFS-S (short-term) — a 3 s window. Smoother; its maximum is what some platforms limit ("short-term max −18").
• LUFS-I (integrated) — one number for the whole file, gated: an absolute gate drops silence below −70 LUFS, then a relative gate drops anything more than 10 LU below the average, so a long quiet intro does not drag a loud song down. This is the number every platform target refers to.

Stereo counts both channels: the same tone on two channels reads 3 LU louder than on one. Multichannel files are measured as BS.1770 says: in a 5.1 file left, right and centre count fully, the surround channels are weighted +1.5 dB and the LFE is left out. The line under the stats names the layout ("5.1"). For podcast delivery in mono or stereo, see Mono or stereo?.

LRA (loudness range)

The spread of the short-term loudness over the file, in LU, between the 10th and 95th percentile after gating (EBU Tech 3342). A podcast conversation sits around 3 to 6 LU; a dynamic live recording can reach 15 or more. High LRA is not a fault — it tells you the quiet parts are far below the loud ones, which matters for listening in a car.

RMS and noise floor (audiobooks)

Audible and ACX do not use LUFS. Their spec is plain RMS between −23 and −18 dB (0 dB is a full-scale square wave), peaks no higher than −3 dBFS, and a noise floor below −60 dB RMS. LoudSense reports the RMS of the whole file and estimates the noise floor from the quietest audible 100 ms blocks — the fifth percentile, which stands in for "the RMS of the room tone". A noise floor problem is reported before any level advice, because gain cannot fix it.

The health checks

• Channel balance — RMS of left against right in dB. More than 1.5 dB usually means a pan or a one-sided microphone.
• Correlation — how alike the channels are. Below zero the channels partly cancel when summed to mono; check phase.
• DC offset — a constant shift of the waveform. It wastes headroom and can click at edits; a 20 Hz high-pass removes it.
• Silence at the start and end, gaps over three seconds inside the programme, and dropouts — runs of exact digital zeros of 20 ms or more, the signature of a buffer underrun or a bad edit.

When there is silence at the start or end, a Trim silence → WAV button cuts it to 0.2 s each side and downloads the result; the original is untouched and the trimmed file is analysed as a new row.

Speech and music

The strip under the panel and the "speech / music" line come from a per-100-ms classifier: speech has strong modulation at syllable rate, little energy below 150 Hz and many zero crossings; music is steadier and bass-heavy. Short runs are smoothed away. Each class gets its own gated LUFS-I, so "speech −18.4, music −13.9" tells you the music sections are 4.5 LU hotter than the voice. It compares the sections, not the bed under the voice, and it is an estimate. Mixed material (sung vocals over a beat, spoken word with a drone) lands wherever the heuristic tips; the numbers are still correct for whatever it labelled.

Why the MP3 clips

Lossy encoding (MP3, AAC, OGG) does not reproduce the waveform exactly. The rebuilt peaks overshoot the original, and the lower the bitrate and the louder the master, the more. On a loud master limited to −1 dBTP we measured +1.7 to +2.8 dBTP after a 128 kbps MP3 encode and up to +1.4 dBTP after 256 kbps AAC — 2.4 to almost 4 dB of overshoot. Quieter, more dynamic material overshoots less. A master with zero clips can come out of the encoder with hundreds, and the decoder in every phone will hit the ceiling on each of them.

The fix is headroom, and checking the encoded file, not only the master: Encode preview encodes it in your browser and measures the result. Fix it lowers its ceiling when you choose MP3 or AAC delivery and checks the encoded result until it holds.

Reading a verdict

Each row compares LUFS-I with the target window and true peak with the ceiling:

• OK — inside the window, peaks under the ceiling.
• Clipped — the file has runs of full-scale samples. Turning it down will not bring the cut-off peaks back; go back to the master before the clip and limit it to the ceiling.
• Too loud — above the window. For a streaming service the consequence line says what it does ("turned down by 2.3 dB"): nothing is lost, but the master is louder than it needs to be. For a broadcast or delivery target it says Out of spec — would be rejected, because a broadcaster sends the file back instead.
• Too quiet — below the window. Spotify (in its normal mode) and Apple Music raise quiet tracks, but only as far as their peaks allow; YouTube, Amazon and Tidal leave them as they are. Either way the file can end up quieter than everything around it. The fix says how much gain, and adds a limiter when the gain would push peaks over the ceiling.
• Peaks too hot — loudness fine, true peak above the ceiling. Limit first, then re-check.

Every row with a fix has a Fix to this button that writes the fixed WAV for that target.

The targets themselves are on Platform targets.