def test_word_to_phoneme(self):
wer_details = WERDetails("../experiments/baseline/wer_details/per_utt",
skip_calculation=True)
self.assertEqual(wer_details.word_to_phoneme("cheese", True),
["CH", "IY", "Z"])
self.assertEqual(wer_details.word_to_phoneme("cheese", False),
["CH", "IY1", "Z"])
def articulatory_barplot(poa_afer: list, label_list: list, reference_label: list, filename: str, phoneme_count: np.ndarray, per: list):
"""
Generates articulatory barplot
:param poa_afer:
:param label_list:
:param reference_label:
:return:
"""
format_dict = {
"width_fig": 3.14,
"height_fig": 3.14 * 0.8, # 0.62
"dpi": 100,
"width": 0.13, # the width of the bars
"fontsize": 7,
"capsize": 1.5
}
num_architectures = len(label_list)
x = np.arange(len(reference_label))
mean_experiments = list()
std_experiments = list()
for experiment in poa_afer:
selected_afer = np.zeros((len(experiment),len(reference_label)))
for i, labelafer in enumerate(experiment):
label, afer = labelafer
idx = [label.index(lab) for lab in reference_label]
afer = np.array(afer)[idx]
selected_afer[i,:] = afer
mean_afer = np.mean(selected_afer,axis=0)
std_afer = np.std(selected_afer,axis=0)
mean_experiments.append(mean_afer)
std_experiments.append(std_afer)
#width_logic = np.linspace(start=-format_dict["width"]*2,stop=format_dict["width"],num=num_architectures)
# Phoeme counts
wer_details = WERDetails("../experiments/baseline/wer_details/per_utt", skip_calculation=True)
phoneme_labels = per[0][0][0]
if "moa" in filename:
af_labels = np.array([wer_details.phoneme_to_moa(phoneme_label) for phoneme_label in phoneme_labels])
if "poa" in filename:
af_labels = np.array([wer_details.phoneme_to_poa(phoneme_label) for phoneme_label in phoneme_labels])
phoneme_counts = np.mean(phoneme_count,axis=0)
af_counts = [round_to_n(np.sum(phoneme_counts[np.where(af_labels == af)]),2) for af in reference_label]
fig = plt.figure(num=None, figsize=(format_dict["width_fig"], format_dict["height_fig"]),
dpi=format_dict["dpi"], facecolor='w', edgecolor='k')
markers = ["v","^","<",">","x"]
for i in range(num_architectures):
# legend_props = {"elinewidth": 0.5}
# plt.bar(x + format_dict["width"]/2 + width_logic[i],
# mean_experiments[i],format_dict["width"],
# label=label_list[i],yerr=std_experiments[i],
# capsize=format_dict["capsize"], error_kw=legend_props)
plt.plot(x, mean_experiments[i], markersize=4, marker=markers[i])
plt.fill_between(x, mean_experiments[i] - std_experiments[i], mean_experiments[i] + std_experiments[i],
alpha=0.2)
r, p = pearsonr(af_counts, mean_experiments[i])
print(label_list[i], " correlation btw data amount and performance", r, "p value", p)
ax = plt.gca()
ax.set_axisbelow(True)
ax.yaxis.grid(color='gray', linestyle='dashed',which="major")
ax.axhline(y=0, color="black",linewidth=0.8)
ax.set_xticks(x)
ax.tick_params(axis='both', which='major', labelsize=format_dict["fontsize"], pad=1)
reference_label_to_plot = [ref + " " + str(int(af_counts[i])) for i,ref in enumerate(reference_label)]
ax.set_xticklabels(reference_label_to_plot, rotation=45, fontsize=format_dict["fontsize"])
plt.ylabel("AFER (%)",fontsize=format_dict["fontsize"])
plt.legend(label_list, fontsize=6,
loc='upper center', bbox_to_anchor=(0.5, +1.32),
fancybox=True, shadow=True, ncol=(num_architectures//2))
plt.xlim([0,len(mean_experiments)])
plt.ylim([20,80])
plt.tight_layout(pad=0)
fig.tight_layout(pad=0)
fig.set_size_inches(format_dict["width_fig"], format_dict["height_fig"])
current_dir = os.path.dirname(__file__)
plt.savefig(os.path.join(current_dir, "figures/" + filename + ".pdf"), bbox_inches='tight', pad_inches=0.005)
print(experiment_folders)
# Purpose of this is to just store pure phoneme count
phoneme_count_per_fold = np.zeros((len(experiment_folders), number_of_phonemes))
for i, experiment in enumerate(experiment_folders):
per_per_experiment, poa_afer_per_experiment, moa_afer_per_experiment,\
poa_cm_per_experiment, moa_cm_per_experiment = [[],[],[],[],[]]
for fold in range(1,6):
t = time.time()
test_folder = [folder for folder in os.listdir(os.path.join("./experiments/",experiment,str(fold)))
if partition in folder]
wer_details = os.path.join("./experiments/",experiment,str(fold),test_folder[0],"wer_details","per_utt")
corpus = WERDetails(wer_details)
per_per_experiment.append(corpus.all_pers())
poa_afer_per_experiment.append(corpus.all_poa_afers())
moa_afer_per_experiment.append(corpus.all_moa_afers())
poa_cm_per_experiment.append(corpus.poa_confusion_matrix())
moa_cm_per_experiment.append(corpus.moa_confusion_matrix())
s = time.time() - t
print("Fold took", s, "seconds")
if i == 0:
phoneme_type, phoneme_counts = np.unique(corpus.all_ref_phonemes, return_counts=True)
phoneme_count_per_fold[fold - 1, :] = phoneme_counts
per.append(per_per_experiment)
poa_afer.append(poa_afer_per_experiment)
def test_poa_to_phonemes(self):
wer_details = WERDetails("../experiments/baseline/wer_details/per_utt",
skip_calculation=True)
self.assertEqual(sorted(wer_details.poa_to_phonemes("Bilabial")),
sorted(["B", "P", "EM", "W", "M"]))
def test_clean_non_words(self):
self.assertEqual(WERDetails.clean_non_words(["aaa", "***", "bbb"]),
["aaa", "bbb"])
def test_arpabet_cleaner(self):
self.assertEqual(WERDetails.arpabet_cleaner("{CH} {IY1} {Z}", True),
["CH", "IY", "Z"])
self.assertEqual(WERDetails.arpabet_cleaner("{CH} {IY1} {Z}", False),
["CH", "IY1", "Z"])
def test_word_in_cmu_dict(self):
wer_details = WERDetails("../experiments/baseline/wer_details/per_utt",
skip_calculation=True)
self.assertEqual(wer_details.word_in_cmu_dict("{asdfhj}"), True)
self.assertEqual(wer_details.word_in_cmu_dict("asdfhj"), False)
def test_phoneme_to_poa(self):
wer_details = WERDetails("../experiments/baseline/wer_details/per_utt",
skip_calculation=True)
self.assertEqual(wer_details.phoneme_to_poa("P"), "Bilabial")
import time
from corpus import WERDetails
from utils import HParam
#partition = "test"
#number_of_phonemes = 40
preprocessing = True
import numpy as np
import pandas as pd
if preprocessing:
config = HParam("configs/dutch.yaml")
wer_details = WERDetails(
"experiments/jasmin_example/scoring_kaldi/wer_details/per_utt",
skip_calculation=False,
config=config)
#phoneme, other = wer_details.all_poa_afers()
t = time.time()
moa_mat = wer_details.moa_confusion_matrix()
s = time.time()
print(s - t, "secs")
poa_mat = wer_details.poa_confusion_matrix()
k = time.time()
print(k - s, "secs")
print(poa_mat)
print()
#df = pd.DataFrame(data=other, index=phoneme)
preprocessing = False
separation = True
if separation:
files = glob("experiments/voicefilter_experiment/*_ss_result.txt")
else:
files = glob("experiments/voicefilter_experiment/*_se_result.txt")
config = HParam("../configs/eng_espnet.yaml")
if preprocessing:
dfs = list()
for file in files:
wer_details = WERDetails(file, skip_calculation=False, config=config)
phoneme, other = wer_details.all_pers()
dfs.append(pd.DataFrame(data=[other[1:]], columns=phoneme[1:], index=[file]))
result = pd.concat(dfs, axis=0, join="outer")
if separation:
result.to_csv("csvs/separation_results_with_clean.csv")
else:
result.to_csv("csvs/enhancement_results_with_clean.csv")
else:
if separation:
df = pd.read_csv("../csvs/separation_results_with_clean.csv")