AudioPrism can feed a freshly created WAV into its separator instead of inserting an MP3 conversion before the stem-splitting stage. The important bit is not the filename extension by itself. What matters is what happened to the audio before Demucs got it.
WAV can hold uncompressed PCM audio, so the separator can receive a decoded waveform without AudioPrism adding another lossy encode first. Easy detail to miss. Stem separation has to work with whatever timing, spectral detail, and transient information survives in the input.
So the useful part of a fresh WAV is much narrower. AudioPrism can avoid taking incoming audio, making an intermediate MP3, decoding that MP3 again, and only then sending the waveform to separation. One lossy generation simply never gets inserted into the chain.
With AudioPrism’s local stem separation workflow, KVR specifically describes linked material going straight into the separator through a fresh WAV rather than a lossy MP3 intermediate. Nothing in that claim means every source started lossless. It means AudioPrism is not deliberately creating another MP3 stage before Demucs does its work.
A compressed source stays compressed in the only sense that matters here. If the audio already came from a low-bitrate MP3 or another lossy delivery format, decoding it to WAV preserves the decoded signal but cannot reconstruct detail removed earlier. Bigger file, same inherited limitations.
A high-bitrate compressed source can still produce perfectly usable stems. Practice parts, rough remixes, reference vocals, and quick MIDI ideas do not suddenly become useless because the source was delivered as MP3. The mistake is treating a later WAV conversion as proof that the underlying input now has lossless-source quality.
Compression also affects machine processing in ways that are not identical to casual listening. Machine perceptual quality under lossy compression evaluated Demucs v3 on music source separation alongside MP3, Opus, and neural audio compression, using an uncompressed MUSDB18-HQ baseline. Its broader result is more nuanced than a simple rule saying every compressed file always ruins machine performance.
Another generation is still worth avoiding. Re-encoding an already compressed signal asks a new encoder to analyze a waveform that has already been altered by an earlier encode, then make another set of keep-or-discard decisions. AudioPrism’s fresh-WAV path removes that avoidable step before separation.
Linked audio needs a more careful expectation. A fresh WAV produced for processing can be a good working file without being a lossless original. Source provenance matters more than the extension sitting in File Explorer.
Sample rate and bit depth labels can be misleading for the same reason. Turning an MP3 into a 24-bit or high-sample-rate WAV gives the decoder more room to store its output, but it does not recover frequencies or fine detail discarded during the original encode. Fancy output settings cannot upgrade the history of the source.
You will usually notice weak input around difficult boundaries rather than across an entire song. Vocal consonants, cymbal tails, piano attacks, distorted guitars, and stacked bright instruments give a separator less room for error because several sounds can occupy similar regions at once. Existing codec damage can make those boundaries even less tidy.
Keep the untouched source beside the separated results instead of repeatedly feeding converted derivatives back into the workflow. If one split sounds rough, rerun from the same starting file or a better source when one becomes available. Comparing from a fixed original also tells you whether the separation changed or the input chain did.
WAV can hold uncompressed PCM audio, so the separator can receive a decoded waveform without AudioPrism adding another lossy encode first. Easy detail to miss. Stem separation has to work with whatever timing, spectral detail, and transient information survives in the input.
Fresh WAV means no extra lossy generation
MP3 saves space by discarding or approximating information according to its encoder and bitrate. Once an MP3 has been decoded into WAV, the discarded material does not magically return. The result is an uncompressed representation of the audio that survived the earlier MP3 encoding.So the useful part of a fresh WAV is much narrower. AudioPrism can avoid taking incoming audio, making an intermediate MP3, decoding that MP3 again, and only then sending the waveform to separation. One lossy generation simply never gets inserted into the chain.
With AudioPrism’s local stem separation workflow, KVR specifically describes linked material going straight into the separator through a fresh WAV rather than a lossy MP3 intermediate. Nothing in that claim means every source started lossless. It means AudioPrism is not deliberately creating another MP3 stage before Demucs does its work.
A compressed source stays compressed in the only sense that matters here. If the audio already came from a low-bitrate MP3 or another lossy delivery format, decoding it to WAV preserves the decoded signal but cannot reconstruct detail removed earlier. Bigger file, same inherited limitations.
Lossy audio can change what a separator receives
Demucs is a waveform-to-waveform separation system, so it works directly from audio samples rather than requiring an MP3 file as its native input. Those samples still reflect every processing step that happened upstream. Removed or altered information cannot help the model draw boundaries between vocals, cymbals, guitars, piano, bass, and everything else competing inside a dense mix.A high-bitrate compressed source can still produce perfectly usable stems. Practice parts, rough remixes, reference vocals, and quick MIDI ideas do not suddenly become useless because the source was delivered as MP3. The mistake is treating a later WAV conversion as proof that the underlying input now has lossless-source quality.
Compression also affects machine processing in ways that are not identical to casual listening. Machine perceptual quality under lossy compression evaluated Demucs v3 on music source separation alongside MP3, Opus, and neural audio compression, using an uncompressed MUSDB18-HQ baseline. Its broader result is more nuanced than a simple rule saying every compressed file always ruins machine performance.
Another generation is still worth avoiding. Re-encoding an already compressed signal asks a new encoder to analyze a waveform that has already been altered by an earlier encode, then make another set of keep-or-discard decisions. AudioPrism’s fresh-WAV path removes that avoidable step before separation.
The cleanest available input still wins
If you already have a proper WAV of the song, there is no sensible reason to manufacture an MP3 copy before separation. Keep the WAV as the source and let AudioPrism work from it. You preserve the information you actually possess instead of shrinking the file first and asking the separator to work from the reduced version.Linked audio needs a more careful expectation. A fresh WAV produced for processing can be a good working file without being a lossless original. Source provenance matters more than the extension sitting in File Explorer.
Sample rate and bit depth labels can be misleading for the same reason. Turning an MP3 into a 24-bit or high-sample-rate WAV gives the decoder more room to store its output, but it does not recover frequencies or fine detail discarded during the original encode. Fancy output settings cannot upgrade the history of the source.
You will usually notice weak input around difficult boundaries rather than across an entire song. Vocal consonants, cymbal tails, piano attacks, distorted guitars, and stacked bright instruments give a separator less room for error because several sounds can occupy similar regions at once. Existing codec damage can make those boundaries even less tidy.
Keep the untouched source beside the separated results instead of repeatedly feeding converted derivatives back into the workflow. If one split sounds rough, rerun from the same starting file or a better source when one becomes available. Comparing from a fixed original also tells you whether the separation changed or the input chain did.
