def analyze(file):
# Feature extraction example
import numpy as np
import librosa
# Load the example clip
y, sr = librosa.load(file)
# Set the hop length; at 22050 Hz, 512 samples ~= 23ms
hop_length = 512
# Separate harmonics and percussives into two waveforms
y_harmonic, y_percussive = librosa.effects.hpss(y)
y_harmonic = abs(y_harmonic)
y_percussive = abs(y_percussive)
# FIXME: this should be sampling rate dependant
max_filter_win_size = 1000
y_harmonic = maximum_filter(y_harmonic, max_filter_win_size)
y_harmonic = gaussian_filter1d(y_harmonic, 1000)
y_percussive = maximum_filter(y_percussive, max_filter_win_size)
y_percussive = gaussian_filter1d(y_percussive, 1000)
# normalize
amax = np.amax(y_harmonic)
y_harmonic = y_harmonic / amax
amax = np.amax(y_percussive)
y_percussive = y_percussive / amax
y_total = y_percussive + y_harmonic
#max_positions = argrelextrema(y_total, np.greater)[0]
max_positions, _ = find_peaks(y_total, distance=150, prominence=0.1)
# open file with wavfile
previous_result = [0] * 84
model = tf.keras.models.load_model("model")
tablature = Tablature()
limits = np.zeros(y_total.shape[0])
for pulse in range(0, len(max_positions) - 1):
time_from = max_positions[pulse]
time_to = max_positions[pulse + 1]
diff = time_to - time_from
DIFF_PERCENTAGE = 0.8
C = np.abs(
librosa.cqt(y[time_from:time_from + int(diff * DIFF_PERCENTAGE)] /
float(65535),
sr=sr,
norm=0,
filter_scale=3))
from_print = (1 / sr) * time_from
to_print = (1 / sr) * (time_from + diff * DIFF_PERCENTAGE)
limits[int(from_print * sr)] = 1
limits[int(to_print * sr)] = 1
# get info
print("note " + str(pulse) + " from: " + str(from_print) + "s to: " +
str(to_print) + "s")
wavfile.write("temp/" + str(pulse) + ".wav", sr,
y[time_from:int(time_from + diff * DIFF_PERCENTAGE)])
# add all samples
result = np.sum(C, axis=1)
amax = np.amax(result)
result = result / amax
new_result = result - previous_result
previous_result = result
# normalize
amax = np.amax(new_result)
# ignore silence pulses
if amax < 0.01:
continue
new_result = new_result / amax
# predict values
prediction = model.predict(np.array([new_result]))
new_result = np.where(prediction[0] == np.amax(prediction[0]))
note_name = new_result[0]
tablature.addNotes(note_name + noteutils.NoteUtils.offsetGuitar)
"""if cur_time != previous_time:
previous_time = cur_time//3
tablature.addTime(str(cur_time))"""
# plot
time = (1 / sr) * len(y_total)
xf = np.linspace(0.0, time, len(y_total))
plt.plot(xf, y_total)
plt.plot(xf, limits)
plt.plot(xf, y)
plt.plot(max_positions / sr, y_total[max_positions], "x")
plt.show()
tablature.print()
