Wiki Home¶

Topics¶

topics/audio-classification-and-representation — Audio tagging, representation learning, audio-language models
topics/classical-techniques — Pre-deep-learning foundations: ASA, CASA, NMF, sinusoidal modeling
topics/datasets — MUSDB18, Slakh2100, GuitarSet, AudioSet, FMA, and more
topics/source-separation — Demucs, Spleeter, BS-RoFormer, AudioSep, MixIT, deep clustering
topics/tools-and-libraries — librosa, pedalboard, music21, MuseScore CLI, impulse response libraries
topics/transcription-and-pitch-estimation — Basic Pitch, MT3, CREPE, TabCNN, Onsets and Frames, SynthTab

[2026-05-19] ingest batch 1 — 10 papers ingested, 11 entities created. First research pass complete.
[2026-05-19] seed — Wiki initialized with 6 topic hubs and 15 concept stubs.

10 sources ingested
42 wiki pages (topics: 6, concepts: 15, entities: 11, comparisons: 0, sources: 10)
6 topics covered
Last updated: 2026-05-19

SOTA progression: Demucs (6.3 dB) → HDemucs (7.7) → HTDemucs (9.2) → BS-RoFormer (9.8)
All tools use fixed stem categories — no bluegrass/string-band specific separation exists
AudioSep offers text-query separation ("separate the banjo") — most novel approach
Open-Unmix is the most hackable for custom fine-tuning

Basic Pitch is the best starting point: instrument-agnostic, pip-installable, handles pitch bends
torchcrepe essential for monophonic F0 tracking of banjo solos
TabCNN is the most relevant architecture — direct audio→tab output
GuitarSet methodology is the template for creating a BanjoSet dataset
No published work on banjo transcription exists — this is an open research gap