From 132dea15fa1c70b06da961c3cc4dffd8f6adb871 Mon Sep 17 00:00:00 2001
From: David Madl <git@abanbytes.eu>
Date: Mon, 27 Apr 2026 11:06:39 +0200
Subject: [PATCH] fix: amplitude cutoff lowpass instead of mean

---
 rhythm.py | 25 +++++++++++++++++++------
 1 file changed, 19 insertions(+), 6 deletions(-)

diff --git a/rhythm.py b/rhythm.py
index f5cfa60..d010712 100644
--- a/rhythm.py
+++ b/rhythm.py
@@ -13,8 +13,8 @@
 
 import numpy as np
 from numpy.fft import fft
-from scipy.signal import fftconvolve
-from scipy.io import wavfile
+
+from hsh_signal.signal import lowpass_fft
 
 def viterbi_highest_frequency_path_vectorized(Scp2, jump_penalty=2.0, use_log_amplitude=True):
     Scp2 = np.asarray(Scp2, dtype=float)
@@ -103,14 +103,21 @@ class BassAnalyzer:
     wavelet_win_sec = 0.175
     k_omega, k_nu = 0.12, 5.0  #: adapt scaling to get 'reasonable' frequency range (for pop bass, e.g. 18..1145 Hz, but that range strongly depends on the actual song's 'pt' shortest interval 'B')
     viterbi_jump_penalty = 5000.0
+    Wp_force = None
+    I_force = None
 
-    def __init__(self, fs, sig):
+    def __init__(self, fs, sig, Wp_force=None):
         """
         :param fs: sampling rate
         :param sig: audio signal normalized to [-1,1]
         """
         self.D = int(self.shift_sec * fs)  #: spectrogram step
-        self.Wp = int(np.round(self.wavelet_win_sec * fs / self.W) * self.W)  # wavelet window - make it an integer multiple of FFT window
+        if self.Wp_force:
+            self.Wp = self.Wp_force
+        elif Wp_force:
+            self.Wp = Wp_force
+        else:
+            self.Wp = int(np.round(self.wavelet_win_sec * fs / self.W) * self.W)  # wavelet window - make it an integer multiple of FFT window
         self.U = self.Wp // self.W  # ratio
 
         self.f = np.pad(sig, (self.W//2, self.W//2-1))  #: signal padded (W-FFT to determine scalogram parameters)
@@ -162,7 +169,10 @@ class BassAnalyzer:
         g = np.abs(Spf)  # (M x W)
         g_bar = np.mean(g, axis=1)  # (M)
         # TODO: check if 'A' needs to be a smooth signal slowly varying over time, not a const.
-        A = np.mean(g_bar)  # amplitude cutoff for pulse train
+        #A = np.mean(g_bar)  # amplitude cutoff for pulse train
+        ip = int(fs)
+        g_bar_l = lowpass_fft(np.pad(g_bar, (ip, ip), mode='edge'), fps=fs, cf=0.5, tw=0.05)[ip:-ip]
+        A = g_bar_l
 
         # compute transitions (pulse train)
         pt = (g_bar > A).astype(int)  # pulse train
@@ -220,7 +230,10 @@ class BassAnalyzer:
         T = (Lp - Wp) / fsp  # un-padded sample count -> time length
         #omega = p * T / B  # width parameter of Gabor wavelet
         #nu = B / (p * T)  # frequency parameter of Gabor wavelet
-        I = int(np.log2(p**2 * T**2 / (delta * B**2)) - 3/2)  # number of octaves
+        if self.I_force:
+            I = self.I_force
+        else:
+            I = int(np.log2(p**2 * T**2 / (delta * B**2)) - 3/2)  # number of octaves
         J = int(256 / I)  # number of voices per octave
 
         r = np.linspace(0, I*J-1, I*J)