dataset is based on kaggle's guitar chords V3

NOTE: Not all chords are included, only:

chords = {
    "Am": [110.00, 164.81, 220.00],  # A, C, E
    "Bb": [116.54, 174.61, 233.08], # Bb, D, F
    "Bdim": [123.47, 174.61, 246.94], # B, D, F
    "C": [130.81, 196.00, 261.63], # C, E, G
    "Dm": [146.83, 220.00, 293.66], # D, F, A
    "Em": [164.81, 246.94, 329.63], # E, G, B
    "F": [174.61, 261.63, 349.23], # F, A, C
    "G": [196.00, 293.66, 392.00] # G, B, D
}

Date: 2024-02-28 using stft to get frequency spectrum

accumulate frequency on time axis

only get frequencies whose magnitude ranks top 6

Date: 2024-03-03 evaluates how closely a given frequency spectrum matches a C major chord. It follows these steps:

1. Generate the C Major Chord Frequency Set: The fundamental frequencies of C (130.81 Hz), G (196.00 Hz), and E (261.63 Hz) are used to create a set of harmonic frequencies by scaling them with powers of 2 (octave shifts).
2. Extract Input Spectrum Data: The detected frequencies and their corresponding magnitudes from an audio signal are provided.
3. Find the Nearest C Chord Frequency: Each input frequency is compared with the C major chord frequency set to find the closest match.
4. Compute Weighted Differences: The absolute frequency difference between each input frequency and its nearest C chord frequency is calculated, then weighted by its magnitude.
5. Summing the Weighted Differences: The final weighted difference sum represents the overall deviation of the input spectrum from the ideal C major chord structure.

NOTE: probably works probably not. Since the frequency & note are not linear relevant, maybe we need to do logarithmic operations on high frequencies, and do exponentiation operations on low frequencies.

trying octave (from C3 to C4) normalization, the effect is good at some previous bad samples (all these samples should be C chord but were previously misclassified:

➝

➝

after test on whole training set, the accuracy for each chord varies:

not sure about the reason but it is definitely related to the structure of different types of chord such as major, minor, diminished.

Date: 2024-03-04: using frequency spectrum top 6 peaks to train the model (PLAN) as it is more readable to human ; - )

Fuck! it doesn't work that much, the accuracy is like 50%, then drop ...

>>> Training Progress:  72%|▋| 721/1000 [00:58<00:21, 13.15it/s, accuracy=0.534
Early stopping triggered at epoch 722.

possible reasons:

1. The frequency spectrum lacks granularity, maybe 6 peaks is too Idealized for neural network.
2. the network itself is not solid enough.