Chroma

Chromagram: CQT magnitudes folded onto pitch classes and peak-normalized per frame, the standard representation for chord and key analysis. Rows are ordered C, C#, D, … for the default twelve classes.

specux.chroma(x, *, sr, hop_length=512, n_chroma=12, bins_per_octave=36,
              n_octaves=7, fmin=None, tuning=0.0, backend="auto")

numpy

import numpy as np

import specux

x = np.random.randn(4 * 22050).astype(np.float32)
C = specux.chroma(x, sr=22050)
C.shape        # (12, 173) = (..., n_chroma, n_frames)

torch

import torch

import specux

xt = torch.randn(8, 4 * 22050, device="cuda", requires_grad=True)
C = specux.chroma(xt, sr=22050)
C.sum().backward()

cupy

import cupy

import specux

xc = cupy.random.standard_normal((8, 4 * 22050), dtype=cupy.float32)
C = specux.chroma(xc, sr=22050)

Parameters

• sr: sample rate in Hz.
• hop_length: frame advance; divisibility as in the CQT.
• n_chroma: pitch classes per octave; must divide bins_per_octave.
• bins_per_octave: CQT bin density under the fold (default 36, three bins per semitone).
• n_octaves: octave span of the underlying CQT (default 7, C1 up).
• fmin: lowest CQT frequency; None means C1, which puts pitch class 0 on C.
• tuning: bin offset in fractions of a bin.
• backend: "auto", "cuda", "cpu", or "metal".

The per-frame peak normalizer is treated as a constant in the backward pass, so gradients flow through the folded magnitudes but not through the normalization itself.

The configured form

specux.Chroma holds the parameters; to_dict() round-trips the configuration. The torch Module for training pipelines is specux.transforms.Chroma.

t = specux.Chroma(sr=22050, n_chroma=12)
C = t(x)                                       # (..., n_chroma, n_frames)
assert specux.Chroma(**t.to_dict()) == t