def read_to_melspecgram():
"""
Reads the raw waveforms and converts them into log-mel spectrograms which
are stored. This is an alternative to read_data() and load_data() to prevent
using too much ram. Trades RAM for disk space.
"""
id_counter = 0
wav_folder = os.path.join(dbc.CRE_DB_PATH, "AudioWAV")
labels_path = os.path.join(dbc.CRE_DB_PATH, "tabulatedVotes.csv")
label_map = get_label_map(labels_path)
for sample_filename in os.listdir(wav_folder):
# Read the sample
sample_path = os.path.join(wav_folder, sample_filename)
wav = load_wav(sample_path)
# Process the sample into a log-mel spectrogram
melspecgram = process_wav(wav, noisy=True)
# Display the spectrogram
display_melspecgram(melspecgram)
# Read the label
sample_name = os.path.splitext(sample_filename)[0]
label = label_map[sample_name]
label = repr_label(label)
# Save the log-mel spectrogram to use later
mel_spec_filename = dbc.CRE_MEL_SPEC_FN.format(id=id_counter,
emo_label=label)
mel_spec_path = os.path.join(dbc.PROCESS_DB_PATH, mel_spec_filename)
np.save(mel_spec_path, melspecgram)
id_counter += 1
def read_data():
"""
Reads the RAVDESS database into np.array.
Sample output:
(array([ 3.0517578e-05, 3.0517578e-05, 3.0517578e-05, ...,
0.0000000e+00, -3.0517578e-05, 0.0000000e+00], dtype=float32), 0)
:return: Tuple of (samples, labels) where the samples are a np.array and the
labels are an array of integers.
"""
samples = []
labels = []
for actor in range(1, NUM_ACTORS + 1):
actor_foldername = "Actor_{:02d}".format(actor)
actor_path = os.path.join(dbc.RAV_DB_PATH, actor_foldername)
print("Processing actor:", actor_foldername)
for sample_filename in os.listdir(actor_path):
# Read the sample and remove the first and last second
sample_path = os.path.join(actor_path, sample_filename)
wav = load_wav(sample_path)
samples.append(remove_first_last_sec(wav, RAV_SR))
# Read the label
label = get_label(
sample_filename, "-", RAV_EMO_INDEX, RAV_EMOTION_MAP)
labels.append(label)
return np.array(samples), np.array(labels)
def read_data():
"""
Reads the CREMA-D database into np.array.
Sample output:
(array([-0.00228882, -0.00204468, -0.00180054, ..., 0. ,
0. , 0. ], dtype=float32), 2)
:return: Tuple of (samples, labels) where the samples are a np.array and the
labels are an array of integers.
"""
samples = []
labels = []
wav_folder = os.path.join(dbc.CRE_DB_PATH, "AudioWAV")
labels_path = os.path.join(dbc.CRE_DB_PATH, "tabulatedVotes.csv")
label_map = get_label_map(labels_path)
for sample_filename in os.listdir(wav_folder):
print("Processing file:", sample_filename)
# Read the sample
sample_path = os.path.join(wav_folder, sample_filename)
samples.append(load_wav(sample_path))
# Read the label
sample_name = os.path.splitext(sample_filename)[0]
label = label_map[sample_name]
labels.append(label)
return np.array(samples), np.array(labels)
def read_to_melspecgram():
"""
Reads the raw waveforms and converts them into log-mel spectrograms which
are stored. This is an alternative to read_data() and load_data() to prevent
using too much ram. Trades RAM for disk space.
"""
id_counter = 0
for actor in range(1, NUM_ACTORS + 1):
actor_foldername = "Actor_{:02d}".format(actor)
actor_path = os.path.join(dbc.RAV_DB_PATH, actor_foldername)
print("Processing actor:", actor_foldername)
for sample_filename in os.listdir(actor_path):
# Read the sample and remove the first and last second
sample_path = os.path.join(actor_path, sample_filename)
wav = remove_first_last_sec(load_wav(sample_path), RAV_SR)
# Process the sample into a log-mel spectrogram
melspecgram = process_wav(wav)
# Display the spectrogram
display_melspecgram(melspecgram)
# Read the label
label = get_label(
sample_filename, "-", RAV_EMO_INDEX, RAV_EMOTION_MAP)
label = repr_label(label)
# Save the log-mel spectrogram to use later
mel_spec_filename = dbc.RAV_MEL_SPEC_FN.format(
id=id_counter, emo_label=label)
mel_spec_path = os.path.join(dbc.PROCESS_DB_PATH, mel_spec_filename)
np.save(mel_spec_path, melspecgram)
id_counter += 1
def read_data():
"""
Reads the TESS database into tensors.
Sample output:
(array([ 3.0517578e-05, 3.0517578e-05, 3.0517578e-05, ...,
0.0000000e+00, -3.0517578e-05, 0.0000000e+00], dtype=float32), 0)
:return: Tuple of (samples, labels) where the samples are a tensor of
varying shape due to varying audio lengths, and the labels are an
array of integers.
"""
samples = []
labels = []
wav_folder = os.path.join(dbc.TES_DB_PATH, "data")
for sample_filename in os.listdir(wav_folder):
print("Processing file:", sample_filename)
# Read the label and if it's empty, then drop the sample
label = get_label(sample_filename, "_", TES_EMO_INDEX, TES_EMOTION_MAP)
if not label:
print("Not using sample:", sample_filename)
continue
# Read the sample
sample_path = os.path.join(wav_folder, sample_filename)
samples.append(load_wav(sample_path))
labels.append(label)
return np.array(samples), np.array(labels)
def read_to_melspecgram():
"""
Reads the raw waveforms and converts them into log-mel spectrograms which
are stored. This is an alternative to read_data() and load_data() to prevent
using too much ram. Trades RAM for disk space.
"""
id_counter = 0
wav_folder = os.path.join(dbc.TES_DB_PATH, "data")
for sample_filename in os.listdir(wav_folder):
print("Processing file:", sample_filename)
sample_path = os.path.join(wav_folder, sample_filename)
# Read the sample
wav = load_wav(sample_path)
# No outlier check because all samples are within 3 stds
# Process the sample into a log-mel spectrogram
melspecgram = process_wav(wav)
# Display the spectrogram
display_melspecgram(melspecgram)
# Read the label
sample_filename = os.path.splitext(sample_filename)[0]
label = get_label(sample_filename, "_", TES_EMO_INDEX, TES_EMOTION_MAP)
label = repr_label(label)
# Save the log-mel spectrogram to use later
mel_spec_filename = dbc.TES_MEL_SPEC_FN.format(
id=id_counter, emo_label=label)
mel_spec_path = os.path.join(dbc.PROCESS_DB_PATH, mel_spec_filename)
np.save(mel_spec_path, melspecgram)
id_counter += 1
def read_to_melspecgram():
"""
Reads the raw waveforms and converts them into log-mel spectrograms which
are stored. This is an alternative to read_data() and load_data() to prevent
using too much ram. Trades RAM for disk space.
"""
id_counter = 0
data_path = os.path.join(dbc.IEM_DB_PATH, "data")
labels_path = os.path.join(dbc.IEM_DB_PATH, "labels")
for sess_num in range(1, NUM_SESS + 1):
sess_foldername = "S{}".format(sess_num)
print("Processing session:", sess_foldername)
sess_data_path = os.path.join(data_path, sess_foldername)
sess_labels_path = os.path.join(labels_path, sess_foldername)
for perform in os.listdir(sess_data_path):
perform_data_path = os.path.join(sess_data_path, perform)
perform_labels_path = os.path.join(
sess_labels_path, "{filename}.txt".format(filename=perform))
print("Processing performance:", perform)
perform_label_map = get_label_map(perform_labels_path)
for sample_filename in os.listdir(perform_data_path):
# Read the sample
sample_path = os.path.join(perform_data_path, sample_filename)
wav = load_wav(sample_path)
# Process the sample into a log-mel spectrogram
melspecgram = process_wav(wav, noisy=True)
# Display the spectrogram
display_melspecgram(melspecgram)
# Get the label
sample_filename = os.path.splitext(sample_filename)[0]
label = repr_label(perform_label_map[sample_filename])
# Save the log-mel spectrogram to use later
mel_spec_filename = dbc.IEM_MEL_SPEC_FN.format(id=id_counter,
emo_label=label)
mel_spec_path = os.path.join(dbc.PROCESS_DB_PATH,
mel_spec_filename)
np.save(mel_spec_path, melspecgram)
id_counter += 1
def read_data():
"""
Reads the IEMOCAP database into np.array.
Sample output:
(array([ 3.0517578e-05, 3.0517578e-05, 3.0517578e-05, ...,
0.0000000e+00, -3.0517578e-05, 0.0000000e+00], dtype=float32), 0)
:return: Tuple of (samples, labels) where the samples are a np.array and the
labels are an array of integers.
"""
samples = []
labels = []
data_path = os.path.join(dbc.IEM_DB_PATH, "data")
labels_path = os.path.join(dbc.IEM_DB_PATH, "labels")
for num_sess in range(1, NUM_SESS + 1):
sess_foldername = "S{}".format(num_sess)
print("Processing session:", sess_foldername)
sess_data_path = os.path.join(data_path, sess_foldername)
sess_labels_path = os.path.join(labels_path, sess_foldername)
for perform in os.listdir(sess_data_path):
perform_data_path = os.path.join(sess_data_path, perform)
perform_labels_path = os.path.join(
sess_labels_path, "{filename}.txt".format(filename=perform))
print("Processing performance:", perform)
perform_label_map = get_label_map(perform_labels_path)
for sample_filename in os.listdir(perform_data_path):
# Read the label and if it's empty, then drop the sample
sample_name = os.path.splitext(sample_filename)[0]
label = perform_label_map[sample_name]
if not label:
print("Not using sample:", sample_filename)
continue
# Read the sample
sample_path = os.path.join(perform_data_path, sample_filename)
samples.append(load_wav(sample_path))
labels.append(label)
return np.array(samples), np.array(labels)