def main(fpath):
# Path to the mp4 file
fpath = glob.glob(fpath + "/*_df.b")[0]
logging.info("Reading file: " + fpath)
with open(fpath, "rb") as f:
vdf = pickle.load(f)
# Find the shape
logging.info("vdf: Number of audio segments:{}".format(vdf.shape[0]))
wer_v = []
wer_wer = []
for val in vdf.iterrows():
ref = val[1]["Reference"]
ds_hyp = val[1]["Deepspeech hypothesis"]
la_hyp = val[1]["Livai hypothesis"]
if not ref:
# Because text.wer throws ZeroDivisionError if ref is null
wer_v.append([1.0, 1.0])
continue
wer_v.append([text.wer(ref, ds_hyp), text.wer(ref, la_hyp)])
wer_wer.append([wer.wer(ref, ds_hyp), wer.wer(ref, la_hyp)])
# Push the wer to data frame for easier calculations
werds_df = pd.DataFrame(wer_v, columns=["WER for DS", "WER for LA"])
# Remove all the values whose WER > 1
werds_df = werds_df[werds_df["WER for DS"] <= 1]
werds_df = werds_df[werds_df["WER for LA"] <= 1]
# Push the wer to data frame for easier calculations
wer_df = pd.DataFrame(wer_wer, columns=["WER for DS", "WER for LA"])
# Remove all the values whose WER > 1
wer_df = wer_df[wer_df["WER for DS"] <= 100]
wer_df = wer_df[wer_df["WER for LA"] <= 100]
# ### Lower WER is better
# Look at these stats
werds_df.describe()
wer_df.describe()
# ### Average of WER
logging.info("Mean of WER using text.py and wer.py are given below: ")
logging.info("text.py WER average for " + fpath + ": " +
str(werds_df.mean()))
logging.info("wer.py WER average for " + fpath + ": " + str(wer_df.mean()))
def convert_to_wer(error):
reference = error[1]
hypothesis = error[2]
wer_v = []
wer_wer = []
for index, ref in enumerate(reference):
try:
wer_v.append([text.wer(ref, hypothesis[index])])
except ZeroDivisionError:
wer_v.append([1.0])
try:
wer_wer.append([wer.wer(ref, hypothesis[index])])
except ZeroDivisionError:
wer_wer.append([100])
# Push the wer to data frame for easier calculations
col_name = "WER for DS"
werds_df = pd.DataFrame(wer_v, columns=[col_name])
# Remove all the values whose WER > 1
werds_df = werds_df[werds_df[col_name] <= 1]
# Push the wer to data frame for easier calculations
wer_df = pd.DataFrame(wer_wer, columns=[col_name])
# Remove all the values whose WER > 1
wer_df = wer_df[wer_df[col_name] <= 100]
# ### Lower WER is better
# Look at these stats
print("{} : {} : {}".format(error[0], werds_df["WER for DS"].mean(),
wer_df["WER for DS"].mean()))
def convert_to_wer(fpath, model_type):
foldername = os.path.basename(fpath)
fpath = os.path.join(fpath, "output_df.b")
# Read the bianry file with lists
logging.info("Reading file: " + fpath)
with open(fpath, "rb") as f:
text_list = pickle.load(f)
# Find the shape
logging.info("Number of audio segments: {}".format(len(text_list[0])))
reference = text_list[0]
if model_type == "la":
hypothesis = text_list[2]
else:
hypothesis = text_list[1]
wer_v = []
wer_wer = []
for index, ref in enumerate(reference):
if not ref:
# Because text.wer throws ZeroDivisionError if ref is null
wer_v.append([1.0, 1.0])
continue
wer_v.append([text.wer(ref, hypothesis[index])])
wer_wer.append([wer.wer(ref, hypothesis[index])])
# Push the wer to data frame for easier calculations
col_name = "WER for " + model_type
werds_df = pd.DataFrame(wer_v, columns=[col_name])
# Remove all the values whose WER > 1
werds_df = werds_df[werds_df[col_name] <= 1]
# Push the wer to data frame for easier calculations
wer_df = pd.DataFrame(wer_wer, columns=[col_name])
# Remove all the values whose WER > 1
wer_df = wer_df[wer_df[col_name] <= 100]
# ### Lower WER is better
# Look at these stats
print(werds_df.describe())
print(wer_df.describe())
# ### Average of WER
logging.info("Mean of WER using text.py and wer.py are given below: ")
logging.info("text.py WER average for " + fpath + ": " +
str(werds_df.mean()))
logging.info("wer.py WER average for " + fpath + ": " + str(wer_df.mean()))
return [foldername, werds_df.mean(), wer_df.mean()]
#!/usr/bin/env python2
# -*- coding: utf-8 -*-
"""
Created on Tue Jan 2 18:01:10 2018
@author: dalonlobo
"""
from __future__ import absolute_import, division, print_function
import text
import sys
import os
if __name__ == "__main__":
folders = []
fpath = os.path.abspath("temp_playlist")
for root, dirs, files in os.walk(fpath):
folders.append(root)
print(folders[1:], file=sys.stderr)
for folder in folders[1:]:
ref_fpath = glob.glob(os.path.join(fpath,folder) + "/*_ref.txt")[0]
hyp_fpath = glob.glob(os.path.join(fpath,folder) + "/*_hyp.txt")[0]
print(ref_path, hyp_path)
with open(ref_fpath) as f1, open(ref_fpath) as f2:
text.wer(f1.read(), f2.read())
def compare_wer(ref, hyp):
with open(ref, "r") as ref, open(hyp, "r") as hyp:
return text.wer(ref.read(), hyp.read())