parent
bc53f726fe
commit
995436c6f1
@ -1,90 +0,0 @@
|
|||||||
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
#
|
|
||||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
||||||
# you may not use this file except in compliance with the License.
|
|
||||||
# You may obtain a copy of the License at
|
|
||||||
#
|
|
||||||
# http://www.apache.org/licenses/LICENSE-2.0
|
|
||||||
#
|
|
||||||
# Unless required by applicable law or agreed to in writing, software
|
|
||||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
||||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
||||||
# See the License for the specific language governing permissions and
|
|
||||||
# limitations under the License.
|
|
||||||
import collections
|
|
||||||
import json
|
|
||||||
|
|
||||||
from paddle.io import Dataset
|
|
||||||
|
|
||||||
from paddleaudio.backends import load as load_audio
|
|
||||||
from paddleaudio.datasets.dataset import feat_funcs
|
|
||||||
|
|
||||||
|
|
||||||
class AMIDataset(Dataset):
|
|
||||||
"""
|
|
||||||
AMI dataset.
|
|
||||||
"""
|
|
||||||
|
|
||||||
meta_info = collections.namedtuple(
|
|
||||||
'META_INFO', ('id', 'duration', 'wav', 'start', 'stop', 'record_id'))
|
|
||||||
|
|
||||||
def __init__(self, json_file: str, feat_type: str='raw', **kwargs):
|
|
||||||
"""
|
|
||||||
Ags:
|
|
||||||
json_file (:obj:`str`): Data prep JSON file.
|
|
||||||
labels (:obj:`List[int]`): Labels of audio files.
|
|
||||||
feat_type (:obj:`str`, `optional`, defaults to `raw`):
|
|
||||||
It identifies the feature type that user wants to extrace of an audio file.
|
|
||||||
"""
|
|
||||||
if feat_type not in feat_funcs.keys():
|
|
||||||
raise RuntimeError(
|
|
||||||
f"Unknown feat_type: {feat_type}, it must be one in {list(feat_funcs.keys())}"
|
|
||||||
)
|
|
||||||
|
|
||||||
self.json_file = json_file
|
|
||||||
self.feat_type = feat_type
|
|
||||||
self.feat_config = kwargs
|
|
||||||
self._data = self._get_data()
|
|
||||||
super(AMIDataset, self).__init__()
|
|
||||||
|
|
||||||
def _get_data(self):
|
|
||||||
with open(self.json_file, "r") as f:
|
|
||||||
meta_data = json.load(f)
|
|
||||||
data = []
|
|
||||||
for key in meta_data:
|
|
||||||
sub_seg = meta_data[key]["wav"]
|
|
||||||
wav = sub_seg["file"]
|
|
||||||
duration = sub_seg["duration"]
|
|
||||||
start = sub_seg["start"]
|
|
||||||
stop = sub_seg["stop"]
|
|
||||||
rec_id = str(key).rsplit("_", 2)[0]
|
|
||||||
data.append(
|
|
||||||
self.meta_info(
|
|
||||||
str(key),
|
|
||||||
float(duration), wav, int(start), int(stop), str(rec_id)))
|
|
||||||
return data
|
|
||||||
|
|
||||||
def _convert_to_record(self, idx: int):
|
|
||||||
sample = self._data[idx]
|
|
||||||
|
|
||||||
record = {}
|
|
||||||
# To show all fields in a namedtuple: `type(sample)._fields`
|
|
||||||
for field in type(sample)._fields:
|
|
||||||
record[field] = getattr(sample, field)
|
|
||||||
|
|
||||||
waveform, sr = load_audio(record['wav'])
|
|
||||||
waveform = waveform[record['start']:record['stop']]
|
|
||||||
|
|
||||||
feat_func = feat_funcs[self.feat_type]
|
|
||||||
feat = feat_func(
|
|
||||||
waveform, sr=sr, **self.feat_config) if feat_func else waveform
|
|
||||||
|
|
||||||
record.update({'feat': feat})
|
|
||||||
|
|
||||||
return record
|
|
||||||
|
|
||||||
def __getitem__(self, idx):
|
|
||||||
return self._convert_to_record(idx)
|
|
||||||
|
|
||||||
def __len__(self):
|
|
||||||
return len(self._data)
|
|
@ -1,175 +0,0 @@
|
|||||||
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
|
|
||||||
#
|
|
||||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
||||||
# you may not use this file except in compliance with the License.
|
|
||||||
# You may obtain a copy of the License at
|
|
||||||
#
|
|
||||||
# http://www.apache.org/licenses/LICENSE-2.0
|
|
||||||
#
|
|
||||||
# Unless required by applicable law or agreed to in writing, software
|
|
||||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
||||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
||||||
# See the License for the specific language governing permissions and
|
|
||||||
# limitations under the License.
|
|
||||||
"""Calculates Diarization Error Rate (DER) which is the sum of Missed Speaker (MS),
|
|
||||||
False Alarm (FA), and Speaker Error Rate (SER) using md-eval-22.pl from NIST RT Evaluation.
|
|
||||||
|
|
||||||
Credits
|
|
||||||
This code is adapted from https://github.com/speechbrain/speechbrain
|
|
||||||
"""
|
|
||||||
import argparse
|
|
||||||
import os
|
|
||||||
import re
|
|
||||||
import subprocess
|
|
||||||
|
|
||||||
import numpy as np
|
|
||||||
|
|
||||||
FILE_IDS = re.compile(r"(?<=Speaker Diarization for).+(?=\*\*\*)")
|
|
||||||
SCORED_SPEAKER_TIME = re.compile(r"(?<=SCORED SPEAKER TIME =)[\d.]+")
|
|
||||||
MISS_SPEAKER_TIME = re.compile(r"(?<=MISSED SPEAKER TIME =)[\d.]+")
|
|
||||||
FA_SPEAKER_TIME = re.compile(r"(?<=FALARM SPEAKER TIME =)[\d.]+")
|
|
||||||
ERROR_SPEAKER_TIME = re.compile(r"(?<=SPEAKER ERROR TIME =)[\d.]+")
|
|
||||||
|
|
||||||
|
|
||||||
def rectify(arr):
|
|
||||||
"""Corrects corner cases and converts scores into percentage.
|
|
||||||
"""
|
|
||||||
|
|
||||||
# Numerator and denominator both 0.
|
|
||||||
arr[np.isnan(arr)] = 0
|
|
||||||
|
|
||||||
# Numerator > 0, but denominator = 0.
|
|
||||||
arr[np.isinf(arr)] = 1
|
|
||||||
arr *= 100.0
|
|
||||||
|
|
||||||
return arr
|
|
||||||
|
|
||||||
|
|
||||||
def DER(
|
|
||||||
ref_rttm,
|
|
||||||
sys_rttm,
|
|
||||||
ignore_overlap=False,
|
|
||||||
collar=0.25,
|
|
||||||
individual_file_scores=False, ):
|
|
||||||
"""Computes Missed Speaker percentage (MS), False Alarm (FA),
|
|
||||||
Speaker Error Rate (SER), and Diarization Error Rate (DER).
|
|
||||||
|
|
||||||
Arguments
|
|
||||||
---------
|
|
||||||
ref_rttm : str
|
|
||||||
The path of reference/groundtruth RTTM file.
|
|
||||||
sys_rttm : str
|
|
||||||
The path of the system generated RTTM file.
|
|
||||||
individual_file_scores : bool
|
|
||||||
If True, returns scores for each file in order.
|
|
||||||
collar : float
|
|
||||||
Forgiveness collar.
|
|
||||||
ignore_overlap : bool
|
|
||||||
If True, ignores overlapping speech during evaluation.
|
|
||||||
|
|
||||||
Returns
|
|
||||||
-------
|
|
||||||
MS : float array
|
|
||||||
Missed Speech.
|
|
||||||
FA : float array
|
|
||||||
False Alarms.
|
|
||||||
SER : float array
|
|
||||||
Speaker Error Rates.
|
|
||||||
DER : float array
|
|
||||||
Diarization Error Rates.
|
|
||||||
"""
|
|
||||||
|
|
||||||
curr = os.path.abspath(os.path.dirname(__file__))
|
|
||||||
mdEval = os.path.join(curr, "./md-eval.pl")
|
|
||||||
|
|
||||||
cmd = [
|
|
||||||
mdEval,
|
|
||||||
"-af",
|
|
||||||
"-r",
|
|
||||||
ref_rttm,
|
|
||||||
"-s",
|
|
||||||
sys_rttm,
|
|
||||||
"-c",
|
|
||||||
str(collar),
|
|
||||||
]
|
|
||||||
print(cmd)
|
|
||||||
if ignore_overlap:
|
|
||||||
cmd.append("-1")
|
|
||||||
|
|
||||||
try:
|
|
||||||
stdout = subprocess.check_output(cmd, stderr=subprocess.STDOUT)
|
|
||||||
|
|
||||||
except subprocess.CalledProcessError as ex:
|
|
||||||
stdout = ex.output
|
|
||||||
|
|
||||||
else:
|
|
||||||
stdout = stdout.decode("utf-8")
|
|
||||||
|
|
||||||
# Get all recording IDs
|
|
||||||
file_ids = [m.strip() for m in FILE_IDS.findall(stdout)]
|
|
||||||
file_ids = [
|
|
||||||
file_id[2:] if file_id.startswith("f=") else file_id
|
|
||||||
for file_id in file_ids
|
|
||||||
]
|
|
||||||
|
|
||||||
scored_speaker_times = np.array(
|
|
||||||
[float(m) for m in SCORED_SPEAKER_TIME.findall(stdout)])
|
|
||||||
|
|
||||||
miss_speaker_times = np.array(
|
|
||||||
[float(m) for m in MISS_SPEAKER_TIME.findall(stdout)])
|
|
||||||
|
|
||||||
fa_speaker_times = np.array(
|
|
||||||
[float(m) for m in FA_SPEAKER_TIME.findall(stdout)])
|
|
||||||
|
|
||||||
error_speaker_times = np.array(
|
|
||||||
[float(m) for m in ERROR_SPEAKER_TIME.findall(stdout)])
|
|
||||||
|
|
||||||
with np.errstate(invalid="ignore", divide="ignore"):
|
|
||||||
tot_error_times = (
|
|
||||||
miss_speaker_times + fa_speaker_times + error_speaker_times)
|
|
||||||
miss_speaker_frac = miss_speaker_times / scored_speaker_times
|
|
||||||
fa_speaker_frac = fa_speaker_times / scored_speaker_times
|
|
||||||
sers_frac = error_speaker_times / scored_speaker_times
|
|
||||||
ders_frac = tot_error_times / scored_speaker_times
|
|
||||||
|
|
||||||
# Values in percentage of scored_speaker_time
|
|
||||||
miss_speaker = rectify(miss_speaker_frac)
|
|
||||||
fa_speaker = rectify(fa_speaker_frac)
|
|
||||||
sers = rectify(sers_frac)
|
|
||||||
ders = rectify(ders_frac)
|
|
||||||
|
|
||||||
if individual_file_scores:
|
|
||||||
return miss_speaker, fa_speaker, sers, ders
|
|
||||||
else:
|
|
||||||
return miss_speaker[-1], fa_speaker[-1], sers[-1], ders[-1]
|
|
||||||
|
|
||||||
|
|
||||||
def main():
|
|
||||||
parser = argparse.ArgumentParser(description="Compute DER")
|
|
||||||
parser.add_argument(
|
|
||||||
"--ref_rttm",
|
|
||||||
type=str,
|
|
||||||
help="the path of reference/groundtruth RTTM file")
|
|
||||||
parser.add_argument(
|
|
||||||
"--sys_rttm",
|
|
||||||
type=str,
|
|
||||||
help="the path of the system generated RTTM file.")
|
|
||||||
parser.add_argument(
|
|
||||||
"--individual_file_scores",
|
|
||||||
type=bool,
|
|
||||||
help="whether returns scores for each file in order.")
|
|
||||||
parser.add_argument("--collar", type=float, help="forgiveness collar.")
|
|
||||||
parser.add_argument(
|
|
||||||
"--ignore_overlap",
|
|
||||||
type=bool,
|
|
||||||
help="whether ignores overlapping speech during evaluation.")
|
|
||||||
|
|
||||||
args = parser.parse_args()
|
|
||||||
|
|
||||||
Scores = DER(args.ref_rttm, args.sys_rttm, args.ignore_overlap, args.collar,
|
|
||||||
args.individual_file_scores)
|
|
||||||
print(Scores)
|
|
||||||
|
|
||||||
|
|
||||||
if __name__ == "__main__":
|
|
||||||
main()
|
|
Loading…
Reference in new issue