Multi-pitch estimation, multi-pitch streaming, and voice assignment¶
Late/Deep, U-Net-Harm, VoasCNN, VoasCLSTM on Cantoría Dataset¶
This notebook contains the code for the figures and evaluation of Appendix A of the dissertation:
Helena Cuesta. Data-driven Pitch Content Description of Choral Singing Recordings. PhD thesis, Universitat Pompeu Fabra. 2022 (to appear).
We compare the performance of Late/Deep, U-Net-Harm, and VoasCNN, proposed models for multi-pitch estimation, multi-pitch streaming, and voice assignment, applied to four-part vocal ensembles.
For this demo, we consider the song Hoy comamos y bebamos from CantorÃa Dataset.
In [1]:
import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
import librosa
import librosa.display
import mir_eval
import os
from IPython.display import Audio, display
In [2]:
FS = 22050
HOPSIZE = 256
F_MIN = 32.7
NUM_FEATURES = 360
BINS_PER_OCTAVE = 60
NUM_OCTAVES = 6
OVER_SAMPLE = 5
def mf0_to_plot_array(times, freqs):
'''Function written by R. Bittner
'''
plot_times = []
plot_freqs = []
for t, freq in zip(times, freqs):
for f in freq:
plot_times.append(t)
plot_freqs.append(f)
return plot_times, plot_freqs
def compute_cqt(x):
cqt = np.abs(librosa.cqt(
y=x, sr=FS, hop_length=HOPSIZE, fmin=F_MIN, n_bins=NUM_FEATURES,
bins_per_octave=BINS_PER_OCTAVE
))
cqt_db = librosa.amplitude_to_db(cqt)
cqt_db = (cqt_db - cqt_db.min()) / (cqt_db.max() - cqt_db.min())
return cqt_db
def compute_hcqt(x):
cqtm = []
for h in [1,2,3,4,5]:
C = librosa.cqt(y=x, sr=FS, hop_length=HOPSIZE,
fmin=F_MIN * h,
n_bins=(NUM_OCTAVES * OVER_SAMPLE * 12),
bins_per_octave=BINS_PER_OCTAVE)
C, P = librosa.magphase(C)
C = librosa.amplitude_to_db(C)
C = (C - C.min()) / (C.max() - C.min())
cqtm.append(C)
cqtm = np.asarray(cqtm).astype(np.float32)
cqtm = np.moveaxis(cqtm, 0, -1)
return cqtm
def get_freq_grid():
"""Get the hcqt frequency grid
"""
freq_grid = librosa.cqt_frequencies(
NUM_OCTAVES * 12 * OVER_SAMPLE, F_MIN, bins_per_octave=BINS_PER_OCTAVE)
return freq_grid
def get_time_grid(n_time_frames):
"""Get the hcqt time grid
"""
time_grid = librosa.core.frames_to_time(
range(n_time_frames), sr=FS, hop_length=HOPSIZE
)
return time_grid
def mpe_to_plot(est_times, est_freqs):
times = []
freqs = []
for t, fqs in zip(est_times, est_freqs):
for f in fqs:
times.append(t)
freqs.append(f)
return np.array(times), np.array(freqs)
def plot_grids(n_frames, sr=22050):
freq_grid = librosa.cqt_frequencies(
n_bins=360,
fmin=32.7,
bins_per_octave=60
)
time_grid = librosa.frames_to_time(
frames = np.arange(n_frames),
sr=sr,
hop_length=256
)
return freq_grid, time_grid
def mono_prep(times, freqs):
freqs = list(freqs)
for i, (tms, fqs) in enumerate(zip(times, freqs)):
if fqs <= 0:
freqs[i] = np.array([])
else:
freqs[i] = np.array([fqs])
return times, freqs
Input song: Hoy comamos y bebamos¶
Cantoria_HCB_Mix.wav
In [3]:
y, sr = librosa.load("./data/Audio/Cantoria_HCB_Mix.wav", sr=44100)
display(Audio(y, rate=sr))
Input features viz¶
In [4]:
cqt = compute_cqt(y)
hcqt = compute_hcqt(y)
In [5]:
fig, ax = plt.subplots(2, 1, figsize=(15,7))
img = librosa.display.specshow(cqt, sr=FS, hop_length=HOPSIZE, x_axis="time",
y_axis="cqt_hz", bins_per_octave=BINS_PER_OCTAVE, ax=ax[0])
ax[0].tick_params(axis='both', labelsize=15)
ax[0].set_ylabel(ylabel="Frequency (Hz)", fontsize=20)
ax[0].set_xlabel(xlabel="")
img = librosa.display.specshow(cqt, sr=FS, hop_length=HOPSIZE, x_axis="time",
y_axis="cqt_hz", bins_per_octave=BINS_PER_OCTAVE, ax=ax[1])
ax[1].tick_params(axis='both', labelsize=15)
ax[1].set_ylabel(ylabel="Frequency (Hz)", fontsize=20)
ax[1].set_xlabel(xlabel="Time (sec)", fontsize=20)
ax[1].set_xlim([5, 15])
plt.tight_layout()
plt.savefig("HCB_input_cqt.png", dpi=200)
In [6]:
fig, ax = plt.subplots(2, 2, figsize=(15,7))
img = librosa.display.specshow(hcqt[:,:,0], sr=FS, hop_length=HOPSIZE, x_axis="time",
y_axis="cqt_hz", bins_per_octave=BINS_PER_OCTAVE, ax=ax[0,0])
ax[0,0].tick_params(axis='both', labelsize=15)
ax[0,0].set_ylabel(ylabel="Frequency (Hz)", fontsize=20)
ax[0,0].set_xlabel(xlabel="")
ax[0,0].set_title("|H[1]|", fontsize=20)
img = librosa.display.specshow(hcqt[:,:,0], sr=FS, hop_length=HOPSIZE, x_axis="time",
y_axis="cqt_hz", bins_per_octave=BINS_PER_OCTAVE, ax=ax[1,0])
ax[1,0].tick_params(axis='both', labelsize=15)
ax[1,0].set_ylabel(ylabel="Frequency (Hz)", fontsize=20)
ax[1,0].set_xlabel(xlabel="Time (sec)", fontsize=20)
ax[1,0].set_xlim([5, 15])
img = librosa.display.specshow(hcqt[:,:,2], sr=FS, hop_length=HOPSIZE, x_axis="time",
y_axis="cqt_hz", fmin=32.7*3, bins_per_octave=BINS_PER_OCTAVE, ax=ax[0,1])
ax[0,1].tick_params(axis='both', labelsize=15)
ax[0,1].set_ylabel(ylabel="Frequency (Hz)", fontsize=20)
ax[0,1].set_xlabel(xlabel="")
ax[0,1].set_title("|H[3]|", fontsize=20)
img = librosa.display.specshow(hcqt[:,:,2], sr=FS, hop_length=HOPSIZE, x_axis="time",
y_axis="cqt_hz", fmin=32.7*3, bins_per_octave=BINS_PER_OCTAVE, ax=ax[1,1])
ax[1,1].tick_params(axis='both', labelsize=15)
ax[1,1].set_ylabel(ylabel="Frequency (Hz)", fontsize=20)
ax[1,1].set_xlabel(xlabel="Time (sec)", fontsize=20)
ax[1,1].set_xlim([5, 15])
plt.tight_layout()
plt.savefig("HCB_input_hcqt.png", dpi=200)
Reference F0 labels¶
We did some manual corrections to the F0 trajectories, automatically extracted with pYIN.
In [7]:
##Â load f0 labels
ref_soprano = pd.read_csv("./data/F0_manual/Cantoria_HCB_S.csv", header=None).values
ref_alto = pd.read_csv("./data/F0_manual/Cantoria_HCB_A.csv", header=None).values
ref_tenor = pd.read_csv("./data/F0_manual/Cantoria_HCB_T.csv", header=None).values
ref_bass = pd.read_csv("./data/F0_manual/Cantoria_HCB_B.csv", header=None).values
fig, ax = plt.subplots(1, 1, figsize=(15, 7))
ax.plot(ref_soprano[:, 0], ref_soprano[:, 1], ".", markersize=5, color="tab:orange", label="Soprano")
ax.plot(ref_alto[:, 0], ref_alto[:, 1], ".", markersize=5, color="tab:green", label="Alto")
ax.plot(ref_tenor[:, 0], ref_tenor[:, 1], ".", markersize=5, color="tab:red", label="Tenor")
ax.plot(ref_bass[:, 0], ref_bass[:, 1], ".", markersize=5, color="tab:blue", label="Bass")
ax.set_ylabel("Frequency (Hz)", fontsize=20)
ax.set_xlabel("Time (sec)", fontsize=20)
ax.tick_params(axis="both", labelsize=15)
ax.set_ylim([80, 500])
ax.legend(fontsize=15, ncol=4, markerscale=3)
plt.savefig("HCB_F0_reference_full.png", dpi=200)
fig, ax = plt.subplots(1, 1, figsize=(15, 7))
ax.plot(ref_soprano[:, 0], ref_soprano[:, 1], ".", markersize=5, color="tab:orange", label="Soprano")
ax.plot(ref_alto[:, 0], ref_alto[:, 1], ".", markersize=5, color="tab:green", label="Alto")
ax.plot(ref_tenor[:, 0], ref_tenor[:, 1], ".", markersize=5, color="tab:red", label="Tenor")
ax.plot(ref_bass[:, 0], ref_bass[:, 1], ".", markersize=5, color="tab:blue", label="Bass")
ax.set_ylabel("Frequency (Hz)", fontsize=20)
ax.set_xlabel("Time (sec)", fontsize=20)
ax.tick_params(axis="both", labelsize=15)
ax.set_ylim([80, 500])
ax.legend(fontsize=15, ncol=4, markerscale=3)
ax.set_xlim([5, 15])
plt.savefig("HCB_F0_reference_zoom.png", dpi=200)
Multiple F0 estimation with Late/Deep¶
Figures with the polyphonic pitch salience representation estimated by Late/Deep, and the resulting multi-pitch stream after post-processing, plotted on top of the reference F0 labels.
For better visualization, we zoom the xaxis between 5 and 15 seconds.
In [8]:
## load Late/Deep prediction, previously computed
ld_salience = np.load("./data/LD/Cantoria_HCB_ld_mpe.npy")
freq_grid, time_grid = plot_grids(ld_salience.shape[1])
fig, ax = plt.subplots(1, 1, figsize=(15, 7))
img = librosa.display.specshow(ld_salience, sr=22050, hop_length=256, x_axis="time",
y_axis="cqt_hz", bins_per_octave=60,ax=ax)
ax.tick_params(axis='both', labelsize=15)
ax.set_ylabel(ylabel="Frequency (Hz)", fontsize=20)
ax.set_xlabel(xlabel="Time (sec)", fontsize=20)
ax.set_xlim([5, 15])
plt.tight_layout()
plt.savefig("HCB_salience_ld_zoom.png", dpi=200)
## load post-processed prediction
pred_times_, pred_freqs_ = mir_eval.io.load_ragged_time_series("./data/LD/Cantoria_HCB_ld_mpe.csv")
pred_times, pred_freqs = mpe_to_plot(pred_times_, pred_freqs_)
## plot references
fig, ax = plt.subplots(1, 1, figsize=(15, 7))
ax.plot(ref_soprano[:, 0], ref_soprano[:, 1], ".", markersize=10, color="tab:blue", label="Soprano")
ax.plot(ref_alto[:, 0], ref_alto[:, 1], ".", markersize=10, color="darkgreen", label="Alto")
ax.plot(ref_tenor[:, 0], ref_tenor[:, 1], ".", markersize=10, color="darkred", label="Tenor")
ax.plot(ref_bass[:, 0], ref_bass[:, 1], ".", markersize=10, color="darkorange", label="Bass")
## plot Late/Deep prediction
ax.plot(pred_times, pred_freqs, ".", markersize=6, color="black", label="L/D prediction")
ax.set_ylabel("Frequency (Hz)", fontsize=20)
ax.set_xlabel("Time (sec)", fontsize=20)
ax.tick_params(axis="both", labelsize=15)
ax.set_ylim([80, 500])
ax.legend(fontsize=15, ncol=5, markerscale=3)
ax.set_xlim([5, 15])
plt.tight_layout()
plt.savefig("HCB_F0_ld_zoom.png", dpi=200)
Numerical evaluation¶
using mir_eval's multi-pitch evaluation metrics
In [9]:
## construct mpe reference following mir_eval's requirements
ref_freqs = list(np.vstack([
ref_soprano[:,1],
ref_alto[:, 1],
ref_tenor[:, 1],
ref_bass[:, 1]
]).transpose())
ref_times = ref_soprano[:,0]
for i, (tms, fqs) in enumerate(zip(ref_times, ref_freqs)):
if any(fqs) <= 20:
ref_freqs[i] = np.array([f for f in fqs if f > 0])
ref_freqs = list(ref_freqs)
## review predictions
pred_freqs_ = list(pred_freqs_)
for i, (tms, fqs) in enumerate(zip(pred_times_, pred_freqs_)):
if fqs.size >= 1:
pred_freqs_[i] = np.array([f for f in fqs if f > 20])
metrics = mir_eval.multipitch.evaluate(ref_times, ref_freqs, pred_times_, pred_freqs_)
metrics["F-Score"] = 2*(metrics["Precision"] * metrics["Recall"]) / (metrics["Precision"] + metrics["Recall"])
print("MPE Results:\n")
print("Precision={}".format(metrics["Precision"]))
print("Recall={}".format(metrics["Recall"]))
print("Accuracy={}".format(metrics["Accuracy"]))
print("F-Score={}".format(metrics["F-Score"]))
Multiple F0 streaming with U-Net-Harm¶
Figures with the monophonic pitch salience representation estimated by U-Net-Harm, and the resulting F0 contours after post-processing.
For better visualization, we zoom the xaxis between 5 and 15 seconds.
In [10]:
sop_sal = np.load("./data/Unetharm/Cantoria_HCB_unh_s.npy")
alt_sal = np.load("./data/Unetharm/Cantoria_HCB_unh_a.npy")
ten_sal = np.load("./data/Unetharm/Cantoria_HCB_unh_t.npy")
bass_sal = np.load("./data/Unetharm/Cantoria_HCB_unh_b.npy")
## plot predicted saliences
fig, ax = plt.subplots(2, 2, figsize=(15, 10))
### soprano \hat{Y}_S
img = librosa.display.specshow(sop_sal, sr=22050, hop_length=256, x_axis="time",
y_axis="cqt_hz", bins_per_octave=60,ax=ax[0,0])
ax[0,0].tick_params(axis='both', labelsize=15)
ax[0,0].set_ylabel(ylabel="Frequency (Hz)", fontsize=20)
ax[0,0].tick_params(axis='both', labelsize=15)
ax[0,0].set_xlabel(xlabel="")
### alto \hat{Y}_A
img = librosa.display.specshow(alt_sal, sr=22050, hop_length=256, x_axis="time",
y_axis="cqt_hz", bins_per_octave=60,ax=ax[0,1])
ax[0,1].tick_params(axis='both', labelsize=15)
ax[0,1].set_ylabel(ylabel="")
ax[0,1].tick_params(axis='both', labelsize=15)
ax[0,1].set_xlabel(xlabel="")
### tenor \hat{Y}_T
img = librosa.display.specshow(ten_sal, sr=22050, hop_length=256, x_axis="time",
y_axis="cqt_hz", bins_per_octave=60,ax=ax[1,0])
ax[1,0].tick_params(axis='both', labelsize=15)
ax[1,0].set_ylabel(ylabel="Frequency (Hz)", fontsize=20)
ax[1,0].tick_params(axis='both', labelsize=15)
ax[1,0].set_xlabel(xlabel="Time (sec)", fontsize=20)
### bass \hat{Y}_B
img = librosa.display.specshow(bass_sal, sr=22050, hop_length=256, x_axis="time",
y_axis="cqt_hz", bins_per_octave=60,ax=ax[1,1])
ax[1,1].tick_params(axis='both', labelsize=15)
ax[1,1].set_ylabel(ylabel="")
ax[1,1].tick_params(axis='both', labelsize=15)
ax[1,1].set_xlabel(xlabel="Time (sec)", fontsize=20)
ax[0,0].set_xlim([5, 15])
ax[0,1].set_xlim([5, 15])
ax[1,0].set_xlim([5, 15])
ax[1,1].set_xlim([5, 15])
plt.tight_layout()
plt.savefig("HCB_salience_unetharm_zoom.png", dpi=200)
In [11]:
sop_f0 = pd.read_csv("./data/Unetharm/Cantoria_HCB_unh_s.csv", header=None).values
alt_f0 = pd.read_csv("./data/Unetharm/Cantoria_HCB_unh_a.csv", header=None).values
ten_f0 = pd.read_csv("./data/Unetharm/Cantoria_HCB_unh_t.csv", header=None).values
bass_f0 = pd.read_csv("./data/Unetharm/Cantoria_HCB_unh_b.csv", header=None).values
## plot predicted F0 contours against the reference F0 labels
fig, ax = plt.subplots(1, 1, figsize=(15, 7))
### plot reference
ax.plot(ref_soprano[:, 0], ref_soprano[:, 1], ".", markersize=10, color="darkblue", label="Soprano")
ax.plot(ref_alto[:, 0], ref_alto[:, 1], ".", markersize=10, color="darkgreen", label="Alto")
ax.plot(ref_tenor[:, 0], ref_tenor[:, 1], ".", markersize=10, color="darkred", label="Tenor")
ax.plot(ref_bass[:, 0], ref_bass[:, 1], ".", markersize=10, color="darkorange", label="Bass")
### plot predictions
ax.plot(sop_f0[:, 0], sop_f0[:, 1], ".", markersize=6, color="lightskyblue", label="Pred. Soprano")
ax.plot(alt_f0[:, 0], alt_f0[:, 1], ".", markersize=6, color="yellowgreen", label="Pred. Alto")
ax.plot(ten_f0[:, 0], ten_f0[:, 1], ".", markersize=6, color="red", label="Pred. Tenor")
ax.plot(bass_f0[:, 0], bass_f0[:, 1], ".", markersize=6, color="gold", label="Pred. Bass")
ax.set_ylabel("Frequency (Hz)", fontsize=20)
ax.set_xlabel("Time (sec)", fontsize=20)
ax.tick_params(axis="both", labelsize=15)
ax.set_ylim([80, 500])
ax.legend(fontsize=15, ncol=4, markerscale=3)
ax.set_xlim([5, 15])
plt.tight_layout()
plt.savefig("HCB_F0_unetharm_zoom.png", dpi=200)
Numerical evaluation¶
In [13]:
## soprano eval
### melody extraction metrics
soprano_metrics = mir_eval.melody.evaluate(
ref_soprano[:,0],
ref_soprano[:,1],
sop_f0[:,0],
sop_f0[:,1]
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_soprano[:,0], ref_soprano[:,1])
est_times, est_freqs = mono_prep(sop_f0[:,0], sop_f0[:,1])
soprano_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
soprano_metrics_mpe["F-Score"] = 2*(soprano_metrics_mpe["Precision"] * soprano_metrics_mpe["Recall"]) / \
(soprano_metrics_mpe["Precision"] + soprano_metrics_mpe["Recall"])
print("SOPRANO Results Melody:")
print("RPA={}".format(soprano_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(soprano_metrics["Overall Accuracy"]))
print("\n")
print("SOPRANO Results MPE:")
print("Precision={}".format(soprano_metrics_mpe["Precision"]))
print("Recall={}".format(soprano_metrics_mpe["Recall"]))
print("Accuracy={}".format(soprano_metrics_mpe["Accuracy"]))
print("F-Score={}".format(soprano_metrics_mpe["F-Score"]))
print("\n\n")
## alto eval
### melody extraction metrics
alto_metrics = mir_eval.melody.evaluate(
ref_alto[:,0],
ref_alto[:,1],
alt_f0[:,0],
alt_f0[:,1]
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_alto[:,0], ref_alto[:,1])
est_times, est_freqs = mono_prep(alt_f0[:,0], alt_f0[:,1])
alto_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
alto_metrics_mpe["F-Score"] = 2*(alto_metrics_mpe["Precision"] * alto_metrics_mpe["Recall"]) / \
(alto_metrics_mpe["Precision"] + alto_metrics_mpe["Recall"])
print("ALTO Results Melody:")
print("RPA={}".format(alto_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(alto_metrics["Overall Accuracy"]))
print("\n")
print("ALTO Results MPE:")
print("Precision={}".format(alto_metrics_mpe["Precision"]))
print("Recall={}".format(alto_metrics_mpe["Recall"]))
print("Accuracy={}".format(alto_metrics_mpe["Accuracy"]))
print("F-Score={}".format(alto_metrics_mpe["F-Score"]))
print("\n\n")
## tenor eval
### melody extraction metrics
tenor_metrics = mir_eval.melody.evaluate(
ref_tenor[:,0],
ref_tenor[:,1],
ten_f0[:,0],
ten_f0[:,1]
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_tenor[:,0], ref_tenor[:,1])
est_times, est_freqs = mono_prep(ten_f0[:,0], ten_f0[:,1])
tenor_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
tenor_metrics_mpe["F-Score"] = 2*(tenor_metrics_mpe["Precision"] * tenor_metrics_mpe["Recall"]) / \
(tenor_metrics_mpe["Precision"] + tenor_metrics_mpe["Recall"])
print("TENOR Results Melody:")
print("RPA={}".format(tenor_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(tenor_metrics["Overall Accuracy"]))
print("\n")
print("TENOR Results MPE:")
print("Precision={}".format(tenor_metrics_mpe["Precision"]))
print("Recall={}".format(tenor_metrics_mpe["Recall"]))
print("Accuracy={}".format(tenor_metrics_mpe["Accuracy"]))
print("F-Score={}".format(tenor_metrics_mpe["F-Score"]))
print("\n\n")
## Bass eval
### melody extraction metrics
bass_metrics = mir_eval.melody.evaluate(
ref_bass[:,0],
ref_bass[:,1],
bass_f0[:,0],
bass_f0[:,1]
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_bass[:,0], ref_bass[:,1])
est_times, est_freqs = mono_prep(bass_f0[:,0], bass_f0[:,1])
bass_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
bass_metrics_mpe["F-Score"] = 2*(bass_metrics_mpe["Precision"] * bass_metrics_mpe["Recall"]) / \
(bass_metrics_mpe["Precision"] + bass_metrics_mpe["Recall"])
print("BASS Results Melody:")
print("RPA={}".format(bass_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(bass_metrics["Overall Accuracy"]))
print("\n")
print("BASS Results MPE:")
print("Precision={}".format(bass_metrics_mpe["Precision"]))
print("Recall={}".format(bass_metrics_mpe["Recall"]))
print("Accuracy={}".format(bass_metrics_mpe["Accuracy"]))
print("F-Score={}".format(bass_metrics_mpe["F-Score"]))
In [14]:
sop_sal = np.load("./data/VA/VoasCNN/Cantoria_HCB_ld_mpe_s.npy")
alt_sal = np.load("./data/VA/VoasCNN/Cantoria_HCB_ld_mpe_a.npy")
ten_sal = np.load("./data/VA/VoasCNN/Cantoria_HCB_ld_mpe_t.npy")
bass_sal = np.load("./data/VA/VoasCNN/Cantoria_HCB_ld_mpe_b.npy")
## plot predicted saliences
fig, ax = plt.subplots(2, 2, figsize=(15, 10))
### soprano \hat{Y}_S
img = librosa.display.specshow(sop_sal, sr=22050, hop_length=256, x_axis="time",
y_axis="cqt_hz", bins_per_octave=60,ax=ax[0,0])
ax[0,0].tick_params(axis='both', labelsize=15)
ax[0,0].set_ylabel(ylabel="Frequency (Hz)", fontsize=20)
ax[0,0].tick_params(axis='both', labelsize=15)
ax[0,0].set_xlabel(xlabel="")
### alto \hat{Y}_A
img = librosa.display.specshow(alt_sal, sr=22050, hop_length=256, x_axis="time",
y_axis="cqt_hz", bins_per_octave=60,ax=ax[0,1])
ax[0,1].tick_params(axis='both', labelsize=15)
ax[0,1].set_ylabel(ylabel="")
ax[0,1].tick_params(axis='both', labelsize=15)
ax[0,1].set_xlabel(xlabel="")
### tenor \hat{Y}_T
img = librosa.display.specshow(ten_sal, sr=22050, hop_length=256, x_axis="time",
y_axis="cqt_hz", bins_per_octave=60,ax=ax[1,0])
ax[1,0].tick_params(axis='both', labelsize=15)
ax[1,0].set_ylabel(ylabel="Frequency (Hz)", fontsize=20)
ax[1,0].tick_params(axis='both', labelsize=15)
ax[1,0].set_xlabel(xlabel="Time (sec)", fontsize=20)
### bass \hat{Y}_B
img = librosa.display.specshow(bass_sal, sr=22050, hop_length=256, x_axis="time",
y_axis="cqt_hz", bins_per_octave=60,ax=ax[1,1])
ax[1,1].tick_params(axis='both', labelsize=15)
ax[1,1].set_ylabel(ylabel="")
ax[1,1].tick_params(axis='both', labelsize=15)
ax[1,1].set_xlabel(xlabel="Time (sec)", fontsize=20)
ax[0,0].set_xlim([5, 15])
ax[0,1].set_xlim([5, 15])
ax[1,0].set_xlim([5, 15])
ax[1,1].set_xlim([5, 15])
plt.tight_layout()
plt.savefig("HCB_salience_voascnn_zoom.png", dpi=200)
In [15]:
## read VoasCNN output F0s (each column in the csv files corresponds to one voice)
f0s = pd.read_csv("./data/VA/VoasCNN/Cantoria_HCB_ld_mpe.csv", header=None).values
times = f0s[:, 0]
sop_f0 = f0s[:, 1]
alt_f0 = f0s[:, 2]
ten_f0 = f0s[:, 3]
bass_f0 = f0s[:, 4]
## plot predicted F0 contours against the reference F0 labels
fig, ax = plt.subplots(1, 1, figsize=(15, 7))
### plot reference
ax.plot(ref_soprano[:, 0], ref_soprano[:, 1], ".", markersize=10, color="darkblue", label="Soprano")
ax.plot(ref_alto[:, 0], ref_alto[:, 1], ".", markersize=10, color="darkgreen", label="Alto")
ax.plot(ref_tenor[:, 0], ref_tenor[:, 1], ".", markersize=10, color="darkred", label="Tenor")
ax.plot(ref_bass[:, 0], ref_bass[:, 1], ".", markersize=10, color="darkorange", label="Bass")
### plot predictions
ax.plot(times, sop_f0, ".", markersize=6, color="lightskyblue", label="Pred. Soprano")
ax.plot(times, alt_f0, "x", markersize=6, color="yellowgreen", label="Pred. Alto")
ax.plot(times, ten_f0, ".", markersize=6, color="red", label="Pred. Tenor")
ax.plot(times, bass_f0, ".", markersize=6, color="gold", label="Pred. Bass")
ax.set_ylabel("Frequency (Hz)", fontsize=20)
ax.set_xlabel("Time (sec)", fontsize=20)
ax.tick_params(axis="both", labelsize=15)
ax.set_ylim([80, 500])
ax.legend(fontsize=15, ncol=4, markerscale=3)
ax.set_xlim([5, 15])
plt.tight_layout()
plt.savefig("HCB_F0_voascnn_zoom.png", dpi=200)
Numerical evaluation¶
In [16]:
## soprano eval
### melody extraction metrics
soprano_metrics = mir_eval.melody.evaluate(
ref_soprano[:,0],
ref_soprano[:,1],
times,
sop_f0
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_soprano[:,0], ref_soprano[:,1])
est_times, est_freqs = mono_prep(times, sop_f0)
soprano_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
soprano_metrics_mpe["F-Score"] = 2*(soprano_metrics_mpe["Precision"] * soprano_metrics_mpe["Recall"]) / \
(soprano_metrics_mpe["Precision"] + soprano_metrics_mpe["Recall"])
print("SOPRANO Results Melody:")
print("RPA={}".format(soprano_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(soprano_metrics["Overall Accuracy"]))
print("\n")
print("SOPRANO Results MPE:")
print("Precision={}".format(soprano_metrics_mpe["Precision"]))
print("Recall={}".format(soprano_metrics_mpe["Recall"]))
print("Accuracy={}".format(soprano_metrics_mpe["Accuracy"]))
print("F-Score={}".format(soprano_metrics_mpe["F-Score"]))
print("\n\n")
## alto eval
### melody extraction metrics
alto_metrics = mir_eval.melody.evaluate(
ref_alto[:,0],
ref_alto[:,1],
times,
alt_f0
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_alto[:,0], ref_alto[:,1])
est_times, est_freqs = mono_prep(times, alt_f0)
alto_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
alto_metrics_mpe["F-Score"] = 2*(alto_metrics_mpe["Precision"] * alto_metrics_mpe["Recall"]) / \
(alto_metrics_mpe["Precision"] + alto_metrics_mpe["Recall"])
print("ALTO Results Melody:")
print("RPA={}".format(alto_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(alto_metrics["Overall Accuracy"]))
print("\n")
print("ALTO Results MPE:")
print("Precision={}".format(alto_metrics_mpe["Precision"]))
print("Recall={}".format(alto_metrics_mpe["Recall"]))
print("Accuracy={}".format(alto_metrics_mpe["Accuracy"]))
print("F-Score={}".format(alto_metrics_mpe["F-Score"]))
print("\n\n")
## tenor eval
### melody extraction metrics
tenor_metrics = mir_eval.melody.evaluate(
ref_tenor[:,0],
ref_tenor[:,1],
times,
ten_f0
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_tenor[:,0], ref_tenor[:,1])
est_times, est_freqs = mono_prep(times, ten_f0)
tenor_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
tenor_metrics_mpe["F-Score"] = 2*(tenor_metrics_mpe["Precision"] * tenor_metrics_mpe["Recall"]) / \
(tenor_metrics_mpe["Precision"] + tenor_metrics_mpe["Recall"])
print("TENOR Results Melody:")
print("RPA={}".format(tenor_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(tenor_metrics["Overall Accuracy"]))
print("\n")
print("TENOR Results MPE:")
print("Precision={}".format(tenor_metrics_mpe["Precision"]))
print("Recall={}".format(tenor_metrics_mpe["Recall"]))
print("Accuracy={}".format(tenor_metrics_mpe["Accuracy"]))
print("F-Score={}".format(tenor_metrics_mpe["F-Score"]))
print("\n\n")
## Bass eval
### melody extraction metrics
bass_metrics = mir_eval.melody.evaluate(
ref_bass[:,0],
ref_bass[:,1],
times,
bass_f0
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_bass[:,0], ref_bass[:,1])
est_times, est_freqs = mono_prep(times, bass_f0)
bass_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
bass_metrics_mpe["F-Score"] = 2*(bass_metrics_mpe["Precision"] * bass_metrics_mpe["Recall"]) / \
(bass_metrics_mpe["Precision"] + bass_metrics_mpe["Recall"])
print("BASS Results Melody:")
print("RPA={}".format(bass_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(bass_metrics["Overall Accuracy"]))
print("\n")
print("BASS Results MPE:")
print("Precision={}".format(bass_metrics_mpe["Precision"]))
print("Recall={}".format(bass_metrics_mpe["Recall"]))
print("Accuracy={}".format(bass_metrics_mpe["Accuracy"]))
print("F-Score={}".format(bass_metrics_mpe["F-Score"]))
Late/Deep + VoasCLSTM¶
In [17]:
sop_sal = np.load("./data/VA/VoasCLSTM/Cantoria_HCB_ld_mpe_s.npy")
alt_sal = np.load("./data/VA/VoasCLSTM/Cantoria_HCB_ld_mpe_a.npy")
ten_sal = np.load("./data/VA/VoasCLSTM/Cantoria_HCB_ld_mpe_t.npy")
bass_sal = np.load("./data/VA/VoasCLSTM/Cantoria_HCB_ld_mpe_b.npy")
## plot predicted saliences
fig, ax = plt.subplots(2, 2, figsize=(15, 10))
### soprano \hat{Y}_S
img = librosa.display.specshow(sop_sal, sr=22050, hop_length=256, x_axis="time",
y_axis="cqt_hz", bins_per_octave=60,ax=ax[0,0])
ax[0,0].tick_params(axis='both', labelsize=15)
ax[0,0].set_ylabel(ylabel="Frequency (Hz)", fontsize=20)
ax[0,0].tick_params(axis='both', labelsize=15)
ax[0,0].set_xlabel(xlabel="")
### alto \hat{Y}_A
img = librosa.display.specshow(alt_sal, sr=22050, hop_length=256, x_axis="time",
y_axis="cqt_hz", bins_per_octave=60,ax=ax[0,1])
ax[0,1].tick_params(axis='both', labelsize=15)
ax[0,1].set_ylabel(ylabel="")
ax[0,1].tick_params(axis='both', labelsize=15)
ax[0,1].set_xlabel(xlabel="")
### tenor \hat{Y}_T
img = librosa.display.specshow(ten_sal, sr=22050, hop_length=256, x_axis="time",
y_axis="cqt_hz", bins_per_octave=60,ax=ax[1,0])
ax[1,0].tick_params(axis='both', labelsize=15)
ax[1,0].set_ylabel(ylabel="Frequency (Hz)", fontsize=20)
ax[1,0].tick_params(axis='both', labelsize=15)
ax[1,0].set_xlabel(xlabel="Time (sec)", fontsize=20)
### bass \hat{Y}_B
img = librosa.display.specshow(bass_sal, sr=22050, hop_length=256, x_axis="time",
y_axis="cqt_hz", bins_per_octave=60,ax=ax[1,1])
ax[1,1].tick_params(axis='both', labelsize=15)
ax[1,1].set_ylabel(ylabel="")
ax[1,1].tick_params(axis='both', labelsize=15)
ax[1,1].set_xlabel(xlabel="Time (sec)", fontsize=20)
ax[0,0].set_xlim([5, 15])
ax[0,1].set_xlim([5, 15])
ax[1,0].set_xlim([5, 15])
ax[1,1].set_xlim([5, 15])
plt.tight_layout()
plt.savefig("HCB_salience_voasclstm_zoom.png", dpi=200)
In [18]:
## read VoasCNN output F0s (each column in the csv files corresponds to one voice)
f0s = pd.read_csv("./data/VA/VoasCLSTM/Cantoria_HCB_ld_mpe.csv", header=None).values
times = f0s[:, 0]
sop_f0 = f0s[:, 1]
alt_f0 = f0s[:, 2]
ten_f0 = f0s[:, 3]
bass_f0 = f0s[:, 4]
## plot predicted F0 contours against the reference F0 labels
fig, ax = plt.subplots(1, 1, figsize=(15, 7))
### plot reference
ax.plot(ref_soprano[:, 0], ref_soprano[:, 1], ".", markersize=10, color="darkblue", label="Soprano")
ax.plot(ref_alto[:, 0], ref_alto[:, 1], ".", markersize=10, color="darkgreen", label="Alto")
ax.plot(ref_tenor[:, 0], ref_tenor[:, 1], ".", markersize=10, color="darkred", label="Tenor")
ax.plot(ref_bass[:, 0], ref_bass[:, 1], ".", markersize=10, color="darkorange", label="Bass")
### plot predictions
ax.plot(times, sop_f0, ".", markersize=6, color="lightskyblue", label="Pred. Soprano")
ax.plot(times, alt_f0, "x", markersize=6, color="yellowgreen", label="Pred. Alto")
ax.plot(times, ten_f0, ".", markersize=6, color="red", label="Pred. Tenor")
ax.plot(times, bass_f0, ".", markersize=6, color="gold", label="Pred. Bass")
ax.set_ylabel("Frequency (Hz)", fontsize=20)
ax.set_xlabel("Time (sec)", fontsize=20)
ax.tick_params(axis="both", labelsize=15)
ax.set_ylim([80, 500])
ax.legend(fontsize=15, ncol=4, markerscale=3)
ax.set_xlim([5, 15])
plt.tight_layout()
plt.savefig("HCB_F0_voasclstm_zoom.png", dpi=200)
Numerical evaluation¶
In [19]:
## soprano eval
### melody extraction metrics
soprano_metrics = mir_eval.melody.evaluate(
ref_soprano[:,0],
ref_soprano[:,1],
times,
sop_f0
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_soprano[:,0], ref_soprano[:,1])
est_times, est_freqs = mono_prep(times, sop_f0)
soprano_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
soprano_metrics_mpe["F-Score"] = 2*(soprano_metrics_mpe["Precision"] * soprano_metrics_mpe["Recall"]) / \
(soprano_metrics_mpe["Precision"] + soprano_metrics_mpe["Recall"])
print("SOPRANO Results Melody:")
print("RPA={}".format(soprano_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(soprano_metrics["Overall Accuracy"]))
print("\n")
print("SOPRANO Results MPE:")
print("Precision={}".format(soprano_metrics_mpe["Precision"]))
print("Recall={}".format(soprano_metrics_mpe["Recall"]))
print("Accuracy={}".format(soprano_metrics_mpe["Accuracy"]))
print("F-Score={}".format(soprano_metrics_mpe["F-Score"]))
print("\n\n")
## alto eval
### melody extraction metrics
alto_metrics = mir_eval.melody.evaluate(
ref_alto[:,0],
ref_alto[:,1],
times,
alt_f0
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_alto[:,0], ref_alto[:,1])
est_times, est_freqs = mono_prep(times, alt_f0)
alto_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
alto_metrics_mpe["F-Score"] = 2*(alto_metrics_mpe["Precision"] * alto_metrics_mpe["Recall"]) / \
(alto_metrics_mpe["Precision"] + alto_metrics_mpe["Recall"])
print("ALTO Results Melody:")
print("RPA={}".format(alto_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(alto_metrics["Overall Accuracy"]))
print("\n")
print("ALTO Results MPE:")
print("Precision={}".format(alto_metrics_mpe["Precision"]))
print("Recall={}".format(alto_metrics_mpe["Recall"]))
print("Accuracy={}".format(alto_metrics_mpe["Accuracy"]))
print("F-Score={}".format(alto_metrics_mpe["F-Score"]))
print("\n\n")
## tenor eval
### melody extraction metrics
tenor_metrics = mir_eval.melody.evaluate(
ref_tenor[:,0],
ref_tenor[:,1],
times,
ten_f0
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_tenor[:,0], ref_tenor[:,1])
est_times, est_freqs = mono_prep(times, ten_f0)
tenor_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
tenor_metrics_mpe["F-Score"] = 2*(tenor_metrics_mpe["Precision"] * tenor_metrics_mpe["Recall"]) / \
(tenor_metrics_mpe["Precision"] + tenor_metrics_mpe["Recall"])
print("TENOR Results Melody:")
print("RPA={}".format(tenor_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(tenor_metrics["Overall Accuracy"]))
print("\n")
print("TENOR Results MPE:")
print("Precision={}".format(tenor_metrics_mpe["Precision"]))
print("Recall={}".format(tenor_metrics_mpe["Recall"]))
print("Accuracy={}".format(tenor_metrics_mpe["Accuracy"]))
print("F-Score={}".format(tenor_metrics_mpe["F-Score"]))
print("\n\n")
## Bass eval
### melody extraction metrics
bass_metrics = mir_eval.melody.evaluate(
ref_bass[:,0],
ref_bass[:,1],
times,
bass_f0
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_bass[:,0], ref_bass[:,1])
est_times, est_freqs = mono_prep(times, bass_f0)
bass_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
bass_metrics_mpe["F-Score"] = 2*(bass_metrics_mpe["Precision"] * bass_metrics_mpe["Recall"]) / \
(bass_metrics_mpe["Precision"] + bass_metrics_mpe["Recall"])
print("BASS Results Melody:")
print("RPA={}".format(bass_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(bass_metrics["Overall Accuracy"]))
print("\n")
print("BASS Results MPE:")
print("Precision={}".format(bass_metrics_mpe["Precision"]))
print("Recall={}".format(bass_metrics_mpe["Recall"]))
print("Accuracy={}".format(bass_metrics_mpe["Accuracy"]))
print("F-Score={}".format(bass_metrics_mpe["F-Score"]))
Post-processing VoasCNN's output with Viterbi decoding¶
In [33]:
sop_sal = np.load("./data/VA/VoasCNN/Cantoria_HCB_ld_mpe_s.npy")
alt_sal = np.load("./data/VA/VoasCNN/Cantoria_HCB_ld_mpe_a.npy")
ten_sal = np.load("./data/VA/VoasCNN/Cantoria_HCB_ld_mpe_t.npy")
bass_sal = np.load("./data/VA/VoasCNN/Cantoria_HCB_ld_mpe_b.npy")
In [43]:
## read VoasCNN output F0s (each column in the csv files corresponds to one voice)
f0s = pd.read_csv("./data/VA/VoasCLSTM/Cantoria_HCB_ld_mpe.csv", header=None).values
times = f0s[:, 0]
sop_f0 = predicted_pitch_sop
alt_f0 = predicted_pitch_alt
ten_f0 = predicted_pitch_ten
bass_f0 = predicted_pitch_bass
## plot predicted F0 contours against the reference F0 labels
fig, ax = plt.subplots(1, 1, figsize=(15, 7))
### plot reference
ax.plot(ref_soprano[:, 0], ref_soprano[:, 1], ".", markersize=10, color="darkblue", label="Soprano")
ax.plot(ref_alto[:, 0], ref_alto[:, 1], ".", markersize=10, color="darkgreen", label="Alto")
ax.plot(ref_tenor[:, 0], ref_tenor[:, 1], ".", markersize=10, color="darkred", label="Tenor")
ax.plot(ref_bass[:, 0], ref_bass[:, 1], ".", markersize=10, color="darkorange", label="Bass")
### plot predictions
ax.plot(times, predicted_pitch_sop, ".", markersize=6, color="lightskyblue", label="Pred. Soprano")
ax.plot(times, predicted_pitch_alt, "x", markersize=6, color="yellowgreen", label="Pred. Alto")
ax.plot(times, predicted_pitch_ten, ".", markersize=6, color="red", label="Pred. Tenor")
ax.plot(times, predicted_pitch_bass, ".", markersize=6, color="gold", label="Pred. Bass")
ax.set_ylabel("Frequency (Hz)", fontsize=20)
ax.set_xlabel("Time (sec)", fontsize=20)
ax.tick_params(axis="both", labelsize=15)
ax.set_ylim([80, 500])
ax.legend(fontsize=15, ncol=4, markerscale=3)
ax.set_xlim([5, 15])
plt.tight_layout()
plt.savefig("HCB_Cantoria_voascnn_viterbi_zoom.png", dpi=200)
In [41]:
## soprano eval
### melody extraction metrics
soprano_metrics = mir_eval.melody.evaluate(
ref_soprano[:,0],
ref_soprano[:,1],
times,
predicted_pitch_sop
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_soprano[:,0], ref_soprano[:,1])
est_times, est_freqs = mono_prep(times, predicted_pitch_sop)
soprano_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
soprano_metrics_mpe["F-Score"] = 2*(soprano_metrics_mpe["Precision"] * soprano_metrics_mpe["Recall"]) / \
(soprano_metrics_mpe["Precision"] + soprano_metrics_mpe["Recall"])
print("SOPRANO Results Melody:")
print("RPA={}".format(soprano_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(soprano_metrics["Overall Accuracy"]))
print("\n")
print("SOPRANO Results MPE:")
print("Precision={}".format(soprano_metrics_mpe["Precision"]))
print("Recall={}".format(soprano_metrics_mpe["Recall"]))
print("Accuracy={}".format(soprano_metrics_mpe["Accuracy"]))
print("F-Score={}".format(soprano_metrics_mpe["F-Score"]))
print("\n\n")
## alto eval
### melody extraction metrics
alto_metrics = mir_eval.melody.evaluate(
ref_alto[:,0],
ref_alto[:,1],
times,
predicted_pitch_alt
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_alto[:,0], ref_alto[:,1])
est_times, est_freqs = mono_prep(times, predicted_pitch_alt)
alto_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
alto_metrics_mpe["F-Score"] = 2*(alto_metrics_mpe["Precision"] * alto_metrics_mpe["Recall"]) / \
(alto_metrics_mpe["Precision"] + alto_metrics_mpe["Recall"])
print("ALTO Results Melody:")
print("RPA={}".format(alto_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(alto_metrics["Overall Accuracy"]))
print("\n")
print("ALTO Results MPE:")
print("Precision={}".format(alto_metrics_mpe["Precision"]))
print("Recall={}".format(alto_metrics_mpe["Recall"]))
print("Accuracy={}".format(alto_metrics_mpe["Accuracy"]))
print("F-Score={}".format(alto_metrics_mpe["F-Score"]))
print("\n\n")
## tenor eval
### melody extraction metrics
tenor_metrics = mir_eval.melody.evaluate(
ref_tenor[:,0],
ref_tenor[:,1],
times,
predicted_pitch_ten
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_tenor[:,0], ref_tenor[:,1])
est_times, est_freqs = mono_prep(times, predicted_pitch_ten)
tenor_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
tenor_metrics_mpe["F-Score"] = 2*(tenor_metrics_mpe["Precision"] * tenor_metrics_mpe["Recall"]) / \
(tenor_metrics_mpe["Precision"] + tenor_metrics_mpe["Recall"])
print("TENOR Results Melody:")
print("RPA={}".format(tenor_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(tenor_metrics["Overall Accuracy"]))
print("\n")
print("TENOR Results MPE:")
print("Precision={}".format(tenor_metrics_mpe["Precision"]))
print("Recall={}".format(tenor_metrics_mpe["Recall"]))
print("Accuracy={}".format(tenor_metrics_mpe["Accuracy"]))
print("F-Score={}".format(tenor_metrics_mpe["F-Score"]))
print("\n\n")
## Bass eval
### melody extraction metrics
bass_metrics = mir_eval.melody.evaluate(
ref_bass[:,0],
ref_bass[:,1],
times,
predicted_pitch_bass
)
### multipitch metrics
ref_times, ref_freqs = mono_prep(ref_bass[:,0], ref_bass[:,1])
est_times, est_freqs = mono_prep(times, predicted_pitch_bass)
bass_metrics_mpe = mir_eval.multipitch.evaluate(ref_times, ref_freqs, est_times, est_freqs)
bass_metrics_mpe["F-Score"] = 2*(bass_metrics_mpe["Precision"] * bass_metrics_mpe["Recall"]) / \
(bass_metrics_mpe["Precision"] + bass_metrics_mpe["Recall"])
print("BASS Results Melody:")
print("RPA={}".format(bass_metrics["Raw Pitch Accuracy"]))
print("Overall Acc={}".format(bass_metrics["Overall Accuracy"]))
print("\n")
print("BASS Results MPE:")
print("Precision={}".format(bass_metrics_mpe["Precision"]))
print("Recall={}".format(bass_metrics_mpe["Recall"]))
print("Accuracy={}".format(bass_metrics_mpe["Accuracy"]))
print("F-Score={}".format(bass_metrics_mpe["F-Score"]))
In [ ]: