Files
pyannote-audio/pyannote/audio/tasks/speaker_tracking/task.py
T
2020-10-12 16:20:50 +02:00

159 lines
5.8 KiB
Python

# MIT License
#
# Copyright (c) 2020 CNRS
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
from pyannote.audio.core.task import TaskSpecification, Problem, Scale, Task
from pyannote.database import Protocol
import math
import random
from pyannote.core import Segment, Timeline, SlidingWindow
from pyannote.core.utils.numpy import one_hot_encoding
class SpeakerTracking(Task):
"""Speaker tracking
Speaker tracking is the process of determining if and when a (previously
enrolled) person's voice can be heard in an audio recording.
Here, it is addressed with the same approach as voice activity detection,
except {"non-speech", "speech"} classes are replaced by {"speaker1", ...,
"speaker_N"} where N is the number of speakers in the training set.
"""
def __init__(
self,
protocol: Protocol,
duration: float = 2.0,
batch_size: int = None,
num_workers: int = 1,
):
super().__init__(
protocol, duration=duration, batch_size=batch_size, num_workers=num_workers
)
# for speaker tracking, task specification depends
# on the data: we do not know in advance which
# speakers should be tracked. therefore, we postpone
# the definition of specifications.
def setup(self, stage=None):
if stage == "fit":
# this is where we load the training set metadata
# to be used later by the train_dataloader.
# here, we simply loop over the training set, remove
# annotated regions shorter than chunk duration, and
# keep track of the reference annotations.
# we also build the list of speakers to be tracked.
self.train, speakers = [], set()
for f in self.protocol.train():
segments = [
segment
for segment in f["annotated"]
if segment.duration > self.duration
]
duration = sum(segment.duration for segment in segments)
self.train.append(
{
"annotated": segments,
"annotation": f["annotation"],
"duration": duration,
"audio": f["audio"],
}
)
speakers.update(f["annotation"].labels())
# now that we now who the speakers are, we can
# define the task specifications.
# note that, since multiple speakers can be active
# at once, the problem is multi-label classification.
self.specifications = TaskSpecification(
problem=Problem.MULTI_LABEL_CLASSIFICATION,
scale=Scale.FRAME,
classes=sorted(speakers),
)
def train__iter__(self):
"""Iterate over training samples
Yields
------
X: (time, channel)
Audio chunks.
y: (frame, num_speakers)
Frame-level targets. Note that frame < time.
`frame` is infered automagically from the
example model output.
"""
random.seed()
while True:
# select one file at random (with probability proportional to its annotated duration)
file, *_ = random.choices(
self.train, weights=[f["duration"] for f in self.train], k=1,
)
# select one annotated region at random (with probability proportional to its duration)
segment, *_ = random.choices(
file["annotated"], weights=[s.duration for s in file["annotated"]], k=1,
)
# select one chunk at random (with uniform distribution)
start_time = random.uniform(segment.start, segment.end - self.duration)
chunk = Segment(start_time, start_time + self.duration)
# extract features
X, _ = self.audio.crop(file, chunk, mode="center", fixed=self.duration)
# TODO | this one_hot_encoding thing needs to be rewritten into pyannote.audio
# TODO | to make sure we always return the same number of frames for the same
# TODO | input duration. we should also support variable-length chunks.
frames = SlidingWindow(
start=chunk.start,
duration=self.frame_duration,
step=self.frame_duration,
)
y = one_hot_encoding(
file["annotation"].crop(chunk),
Timeline([chunk]),
frames,
labels=self.specifications.classes,
mode="center",
).data
yield {"X": X, "y": y}
def train__len__(self):
# Number of training samples in one epoch
duration = sum(file["duration"] for file in self.train)
return math.ceil(duration / self.duration)