Lossless audio checker

Spectral-cutoff detection, measured instead of eyeballed

A "lossless" file traded online is sometimes a 128 kbps MP3 re-encoded to FLAC. The classic tell is a spectrogram with a hard wall around 16 kHz, the Spek screenshot audiophile forums have traded for years. This page turns that manual read into a number: @audio/measure-lossy computes the long-term average spectrum (Welch, via @audio/spectral-ltas) and a max-hold spectrum, finds the sustained drop that marks the true cutoff, and matches it against the published lowpass tables of LAME MP3, AAC, Vorbis and Opus.

What the numbers mean

• Cutoff frequency: the highest frequency the file actually carries. Compare it with the file's own Nyquist, half its sample rate; a gap is the signature of a codec lowpass filter.
• Edge sharpness: how steep the drop is, in dB per octave. Psychoacoustic encoders cut sharply, above 60 dB/octave; microphones and tape roll off gently.
• Spectral holes: short 4-16 kHz gaps per second where a low-bitrate encoder ran out of bits, the classic MP3 swiss-cheese pattern.
• Upsampled: a cutoff sitting at a clean fraction of the sample rate, the 22.05, 24 or 32 kHz half-rates, means a lower-rate source padded back up, not a bitrate-driven filter.

Where this method breaks down

Above roughly 128 to 160 kbps a well-tuned MP3 or AAC often shows no measurable lowpass, so it can read as lossless even though it was transcoded. The reverse happens too: an old tape dub, a narrow-band microphone, or plain speech is naturally band-limited and can look exactly like a lossy cutoff. This is a known limit of spectral-cutoff detection generally, shared by Spek and aucdtect-style guessers, not a defect specific to this page. Treat the verdict as one data point and read the cutoff, sharpness and holes together.