StemDeck splits any track into 6 stems locally, free and Apache-2.0
It wraps Meta's Demucs to run fully offline — no quotas, no uploads. Also fresh: Claude Science opens in beta and Hindi joins the Open ASR Leaderboard.
Copy markdown
Six stems, nothing leaves your box
StemDeck wraps Meta's Demucs `htdemucs_6s` (~170 MB, auto-downloaded once) to split vocals, drums, bass, guitar, piano, and 'other,' auto-detecting NVIDIA CUDA, Apple MPS, or a CPU fallback. It's Apache-2.0 with native macOS/Windows/Linux builds plus Docker and Unraid images, and your audio never touches a server.
The weekend build: a zero-cost mixer stack
Beyond separation you get a browser-based DAW mixer, BPM/key/scale and LUFS detection, click-track generation, and YouTube/SoundCloud import — all local. Fork it into a karaoke app, a sample miner, or a practice-track pipeline without per-minute cloud fees or upload caps.
Claude Science lands in beta
Anthropic's new research workbench ships 60+ skills for genomics, proteomics, and cheminformatics, natively renders 3D protein, genome, and chemical structures, and queries 60+ databases like UniProt and PDB — running locally or over SSH/HPC so sensitive data stays put. It's in beta for Pro, Max, Team, and Enterprise on macOS and Linux, with discounted seats for academic and nonprofit labs.
Hindi cracks the Open ASR Leaderboard
Hugging Face added Hindi — its first Indic language — to the Open ASR Leaderboard via the Monsoon datasets: 4,888 speakers across 428+ districts and nine axes of variation, scored with a new OIWER metric for spelling variants. The lesson for STT builders: models tied on aggregate WER can diverge by region (Voxtral-Mini shifts 1.68 points across zones), so evaluate disaggregated before you ship.