plate iii — raw audio
Muud
A desktop app that hears what a song is — and how it feels.
- signal
- raw audiomel spectrograms
- instrument
- tensorflowcrnn + fuzzy fusion
- role
- solo: datasets, training, app
- stack
- tensorflow/keras · librosa
- year
- 2026
- links
- repo ↗
the reading
excitement
valence 7.4 · arousal 7.4
On DEAM’s 1–9 scale, the one Muud’s regression head learns. Muud fuses the point it hears with the genre it hears: 0.7 × genre similarity + 0.3 × emotion similarity.
drag the disc · or focus it and use ← →
i. problem
Genre labels are flat. “Rock” tells you nothing about whether a track belongs at a funeral or a party — the feeling is the signal that matters for recommendation, and feeling is continuous, not categorical.
ii. signal
Raw audio, decomposed twice. Mel spectrograms feed a 10-class CRNN trained on FMA-medium (temperature-scaled, multi-segment averaged, with hybrid labels when the top two genres land within 0.10). In parallel, a hybrid CNN plus handcrafted features — tempo, spectral centroid, RMS, zero-crossing rate — regresses continuous valence and arousal on the DEAM 1–9 scale.
iii. build
The two heads meet in fuzzy fusion: 0.7 × graded genre similarity (an inter-genre matrix, so blues is near rock, not simply ≠) + 0.3 × emotion similarity, driving Spotify-backed recommendations. A live microphone mode runs a rolling spectrogram with inference every ~7 seconds and temporal smoothing across five windows. And because a model you can't interrogate is a mood ring, an explainability panel shows the actual fusion arithmetic, genre memberships, and a valence-arousal plot on Russell's circumplex.
iv. proof
~68% validation accuracy on 10-way FMA genre classification — an honest number for a hard dataset — with both models trained from scratch, not rented from an API. Every recommendation can show its work.