v0.2.0-beta

app.py

@@ -4,7 +4,7 @@ cqfe_interface = gr.Interface(fn=cqfe,
                               inputs=gr.Audio(type='filepath', format='wav', label='Audio Input File'),
                               outputs=[gr.File(type='file', label='F0 Output Files'),
                                        gr.Plot(label='F0 Estimation Plot')],
-                              title="Choral Quartets F0 Extractor (v0.1.3-alpha)",
+                              title="Choral Quartets F0 Extractor (v0.2.0-beta)",
                               description="An application that uses Multi-Pitch Estimation and Voice Assignment to transform audio files with Choral Quartets recordings into files (CSV, HDF5 and MIDI) containing F0 estimations for each voice (Soprano, Alto, Tenor and Bass). The processing may take a few minutes.")
 
-cqfe_interface.launch()
+cqfe_interface.launch()

app_test.ipynb

@@ -12,7 +12,7 @@
     "                              inputs=gr.Audio(type='filepath', format='wav', label='Audio Input File'),\n",
     "                              outputs=[gr.File(type='file', label='F0 Output Files'),\n",
     "                                       gr.Plot(label='F0 Estimation Plot')],\n",
-    "                              title=\"Choral Quartets F0 Extractor (v0.1.3-alpha)\",\n",
+    "                              title=\"Choral Quartets F0 Extractor (v0.2.0-beta)\",\n",
     "                              description=\"An application that uses Multi-Pitch Estimation and Voice Assignment to transform audio files with Choral Quartets recordings into files (CSV, HDF5 and MIDI) containing F0 estimations for each voice (Soprano, Alto, Tenor and Bass). The processing may take a few minutes.\")\n",
     "\n",
     "cqfe_interface.launch()"

cqfe_utils.py

@@ -8,6 +8,8 @@ import pandas as pd
 from scipy.ndimage import gaussian_filter1d
 from cqfe_models import mask_voas_cnn_v2_model, late_deep_cnn_model
 
+SATB_THRESHOLDS = [0.23, 0.17, 0.15, 0.17]
+
 ############################################################
 
 freqscale = librosa.cqt_frequencies(n_bins=360, fmin=32.7, bins_per_octave=60)
@@ -134,9 +136,9 @@ def song_to_midi(sop, alto, ten, bass):
     freq_matrix = bin_matrix_to_freq(bin_matrix)
 
     mid_sop = create_midi(freq_matrix[0], save_to_file=False, program=52, channel=0)
-    mid_alto = create_midi(freq_matrix[1], save_to_file=False, program=53, channel=1)
-    mid_ten = create_midi(freq_matrix[2], save_to_file=False, program=49, channel=2)
-    mid_bass = create_midi(freq_matrix[3], save_to_file=False, program=50, channel=3)
+    mid_alto = create_midi(freq_matrix[1], save_to_file=False, program=52, channel=1)
+    mid_ten = create_midi(freq_matrix[2], save_to_file=False, program=42, channel=2)
+    mid_bass = create_midi(freq_matrix[3], save_to_file=False, program=52, channel=3)
 
     mid_mix = mido.MidiFile()
     mid_mix.ticks_per_beat=mid_sop.ticks_per_beat
@@ -165,11 +167,14 @@ def song_to_dataframe(sop, alto, ten, bass):
 
 ############################################################
 
-def prediction_postproc(input_array, argmax_and_threshold=True, gaussian_blur=True):
+def prediction_postproc(input_array, argmax_and_threshold=True,
+                                     gaussian_blur=True,
+                                     threshold_value=0):
     prediction = np.moveaxis(input_array, 0, 1).reshape(360, -1)
+    thres_reference = deepcopy(prediction)
     if(argmax_and_threshold):
         prediction = np.argmax(prediction, axis=0)
-        prediction = np.array([i if i <= 357 else 0 for i in prediction])
+        prediction = np.array([prediction[i] if thres_reference[prediction[i], i] >= threshold_value else 0 for i in np.arange(prediction.size)])
         threshold = np.zeros((360, prediction.shape[0]))
         threshold[prediction, np.arange(prediction.size)] = 1
         prediction = threshold
@@ -273,10 +278,10 @@ def get_va_prediction(model, f0_matrix):
         t_pred_result = np.append(t_pred_result, t_pred, axis=0)
         b_pred_result = np.append(b_pred_result, b_pred, axis=0)
 
-    s_pred_result = prediction_postproc(s_pred_result)[:, :f0_matrix.shape[1]]
-    a_pred_result = prediction_postproc(a_pred_result)[:, :f0_matrix.shape[1]]
-    t_pred_result = prediction_postproc(t_pred_result)[:, :f0_matrix.shape[1]]
-    b_pred_result = prediction_postproc(b_pred_result)[:, :f0_matrix.shape[1]]
+    s_pred_result = prediction_postproc(s_pred_result, threshold_value=SATB_THRESHOLDS[0])[:, :f0_matrix.shape[1]]
+    a_pred_result = prediction_postproc(a_pred_result, threshold_value=SATB_THRESHOLDS[1])[:, :f0_matrix.shape[1]]
+    t_pred_result = prediction_postproc(t_pred_result, threshold_value=SATB_THRESHOLDS[2])[:, :f0_matrix.shape[1]]
+    b_pred_result = prediction_postproc(b_pred_result, threshold_value=SATB_THRESHOLDS[3])[:, :f0_matrix.shape[1]]
 
     return s_pred_result, a_pred_result, t_pred_result, b_pred_result
 

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "Choral-Quartets-F0-Extractor"
-version = "0.1.3-alpha"
+version = "0.2.0-beta"
 description = "An application that uses Multi-Pitch Estimation and Voice Assignment to transform audio files with Choral Quartets recordings into files (CSV, HDF5 and MIDI) containing F0 estimations for each voice (Soprano, Alto, Tenor and Bass)."
 authors = ["André Paiva (Xornotor) <[email protected]>"]
 license = "cc"