Add Automatic Music Transcription/MIDI generation

This release adds Automatic Music Transcription and MIDI generation to the website for harmonic/pitched instruments!

Add Demucs v3 (hdemucs_mmi) support

Demucs v3 Hybrid model (hdemucs_mmi weights) is now available on the free site.

This model is faster than the v4 transformer models, and the performance is comparable (and better in some situations).

Pro site released

I've been working on new models after the conclusion of the Sound Demixing Challenge 2023

Ensemble models dominated, with mvsep using an interesting technique of inference augmentation with the inverted waveform:

stems = model_1(waveform)
inverse_stems = model_1(waveform * -1.0) * -1.0

I'm launching the pro site with special performing ensembles of demucs:

Demucs is now 25% faster

It pays to read the docs sometimes!

Enabling WASM's SSE4.2 intrinsics translation by supplying the -msse4.2 flag alongside -msimd128 led to a significant speed boost, shaving off 25% of the demixing time for a song.

Demucs multi-threading + 6-source (guitar, piano stems)

This release adds some great new features:

• Multi-threaded Demucs for much faster processing!
  • MAX MEMORY setting (use 16 GB or 32 GB to run Demucs faster with multiple parallel threads/workers)
  • Each worker consumes 4 GB of memory
• Make Demucs 4-source model the default; drop/deprecate Open-Unmix/UMX (you can still check the same umx.cpp code here)
• Add Demucs 6-source model, with experimental piano and guitar stems! 🎹 🎸

Demucs v4 is now available!

The Demucs v4 hybrid transformer model is now running in WASM on this site!

Read the announcement post or visit the site to try it out!

The underlying C++ code comes from https://github.com/sevagh/demucs.cpp

New release: full UMX-L performance

Implemented:

• Wiener filtering
• Segmented inference (from Demucs)
• Streaming LSTM (custom architecture)
• Batch uploading
• Dev logs embedded on the frontend