Agentic Music Transcription Monitor

This week's most relevant papers combining Automatic Music Transcription with LLMs, agents, or foundation models. Clickable arXiv links, code repos, and a Copy Prompt button to re-run the agent.

Recent activity across Demucs, Basic Pitch, MT3, music21, LangGraph, crewAI. Stars, latest releases, breaking changes. 1 repo gone (omnizart) and 1 dormant (basic-pitch).

CC-licensed test excerpts (30-second) and eval sets (50-100 pieces) for the pipeline. License verification on every source — NC flags called out.

Every clickable link from every report, in one searchable page. Filter by title, source, or URL. Handy when you want to drop a batch into your browser tabs.

Upload a solo instrument audio (or use a built-in sample) and the browser does monophonic pitch detection. Generates a multi-instrument description (key, range, suggested arrangement) from the detected notes.

Full AMT standalone: drop audio or hold a mic, get FFT+HPS pitch detection, VexFlow notation, piano roll, and exports to SVG / PDF / MIDI / MusicXML / WAV segments. Dark + light theme.

Two tools in one: the audio-production resource cards (Ableton Link, RTP-MIDI, BlackHole, AudioGridder, AUM, REW + ResoScan) and a Session Configurator that turns your devices, DAW, iOS apps, and audio interface into six copy-pasteable setup blocks — sync config, audio routing, MIDI routing, OSC mapping, latency budget, and dual-destination recording. Vanilla JS, copy buttons.

Two tools in one page. Part A: the cycle-4 AMT resource cards (ACE-Step v1.5, Libretto, Rubato, Stable Audio 3.0, V2N) with cycle-3 winners carried forward. Part B: a Long Horizon Prompt generator — type a creative idea, get nine production-ready prompts: horizontal / vertical / square video, audio, music, social post, plus three new publishing outputs (lead sheet, engraving spec, MEI prep). Vanilla JS, copy buttons on every output.

Stage 2.5 of the pipeline. Take a melody (text or built-in tune), pick one of the four MIDI-DDSP pretrained timbres (violin / flute / sax / trumpet, Apache 2.0), dial in the expression, hit Play — the browser uses the Web Audio API to synthesise it locally and plays through your speakers. Export WAV downloads the rendered audio. No fetch, no model download, no API. The real MIDI-DDSP neural synth path lives in setup/ddsp_transfer.py.