The Stem Separator just got a major upgrade — v5 — with ten new features that turn it from a "separate and download" tool into a full production workstation. Everything runs on the same local RTX 2070 (no cloud, no accounts, no per-track fees).

What's new

• 🎚️ Stem Mix & Export — set per-stem volumes, mute tracks, preview the mix live in your browser, then export the mixed result as WAV or MP3 (with optional -14 LUFS normalization for streaming platforms).
• ⚖️ A/B Compare — play any two sources (original vs. any stem) in perfect sync with a blend slider and overlaid waveforms. Press C to toggle.
• 🧠 10-Model Picker — beyond the default 6-stem split, choose male/female vocal splits, chorus split, de-reverb (dry/wet), or the revival variants. Models download on first use and are cached locally.
• 📦 Batch Upload — drop in a whole album. Files queue on the GPU (one at a time), each with its own progress, position, and cancel button.
• ✕ Cancel Anytime — stop a running job mid-separation and the GPU is freed immediately for the next one.
• 📌 Persistent Library — pin results to keep them forever (exempt from the 24-hour cleanup), and share any result with a single link: stem-separator/?task=…
• ♩ BPM & Key Detection — every separation now reports tempo and key, with beat-aligned export (trim to the first 4 or 8 bars, DAW-ready).
• 🎛️ Pro Export Formats — WAV 16/24-bit, FLAC, MP3 at 128–320kbps, 48kHz, and two-pass LUFS loudness normalization. One dropdown per stem.
• 🎹 MIDI Piano Roll — transcribe any stem to MIDI and view it in an interactive piano roll (hover for note info, click to hear it, download the .mid).
• 📊 Live Dashboard — the header now shows queue depth, GPU temperature/utilization, and a live ETA for the running job.

Under the hood

The backend was rebuilt around a cancellable inference loop (verified bit-identical to the original), a serialized GPU queue, an LRU model cache (max 2 models in VRAM), and a new /dashboard API. The frontend is still a single dependency-free HTML file. All 10 features passed a full server regression suite plus a 14-step headless-browser E2E — zero console errors.

One item is pending: live MIDI transcription needs a free HuggingFace token (the MuScriptor weights are license-gated). The piano-roll UI, note parsing, and all-stems MIDI export are already built and tested — it goes live the moment the token is added.

Try it: Stem Separator