Repository navigation
What's new in 0.2.0
Key and tempo, read automatically
Every separation now reports the track's tempo and its musical key, in both
note and Camelot form, so you can beatmatch and mix harmonically without a
second tool. Tempo has ×2 and ÷2 buttons, because which metrical level a
track "is" is genuinely ambiguous — even librosa answers 129 or 64 for the
same track depending on where you start it.
Unusually, this runs on the separated stems rather than the mix: tempo
comes from the isolated drums, key from bass + other with drums and vocals
removed. Chroma is very sensitive to broadband percussion, so having the
stems is a real advantage no tool analysing a finished track can match.
Measured against BPM and key tags written by Serato and rekordbox on a real
DJ library:
| result | |
|---|---|
| Tempo, exact | 17 of 18 tagged tracks (94%) |
| Tempo, within an octave | 18 of 18 (100%) |
| Key, exact | 39% — librosa's own CQT chroma scores 35% |
| Key, exact or harmonically adjacent | 68% — librosa scores 56% |
Key detection is the weaker half and the app says so rather than pretending:
confidence is shown, and below 0.6 it is labelled a guess, because at that
level it was right only about one time in five.
Play all four stems together
The results pane now plays every stem at once, with mute, solo and a level
fader on each, plus a scrub bar and spacebar transport — so you can check a
mix without leaving the app. Everything stays sample-locked no matter how
much you mute, solo and scrub.
Faster, with bit-identical output
Asking for fewer than four stems is now dramatically quicker on the Maximum
quality setting. htdemucs_ft is four models with one-hot weights — model 0
is fine-tuned for drums, model 1 for bass, and so on — so the app was running
models whose entire contribution it then multiplied by zero and discarded.
| job | before | after |
|---|---|---|
| Vocals only, processor | 52.6s | 9.2s |
| Drums + bass, processor | 52.6s | 20.1s |
| Vocals only, NVIDIA GPU | — | 1.6s |
This is not a quality trade: the output is byte-for-byte identical, because
the skipped terms were multiplied by 0.0 anyway. Asking for all four stems is
unchanged.
Smaller things
- A file passed on the command line now opens, so Open with → StemSplit
Studio works from Explorer. - The track's length now shows on the file card before you separate. A
partial download looks completely normal until the stems come out short. - Your own logo is in the title bar.
- A file that decodes to no audio now says so clearly instead of failing
obscurely.
Not changed, deliberately
Demucs' shifts setting was measured against MUSDB18's real ground-truth
stems and left switched off: it buys +0.10 dB mean SI-SDR for 40% more
time, which is far below the ~1 dB where a difference becomes audible.
First run: Windows may say "Windows protected your PC" because the app is
not code-signed. Click More info, then Run anyway.