File size: 4,814 Bytes
7934b29 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 |
# Copyright (c) 2020, NVIDIA CORPORATION. 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.
"""
This is a helper script to extract speaker embeddings based on manifest file
Usage:
python extract_speaker_embeddings.py --manifest=/path/to/manifest/file'
--model_path='/path/to/.nemo/file'(optional)
--embedding_dir='/path/to/embedding/directory'
Args:
--manifest: path to manifest file containing audio_file paths for which embeddings need to be extracted
--model_path(optional): path to .nemo speaker verification model file to extract embeddings, if not passed SpeakerNet-M model would
be downloaded from NGC and used to extract embeddings
--embeddings_dir(optional): path to directory where embeddings need to stored default:'./'
"""
import json
import os
import pickle as pkl
from argparse import ArgumentParser
import numpy as np
import torch
from nemo.collections.asr.models.label_models import EncDecSpeakerLabelModel
from nemo.collections.asr.parts.utils.speaker_utils import embedding_normalize
from nemo.utils import logging
def get_embeddings(speaker_model, manifest_file, batch_size=1, embedding_dir='./', device='cuda'):
"""
save embeddings to pickle file
Args:
speaker_model: NeMo <EncDecSpeakerLabel> model
manifest_file: path to the manifest file containing the audio file path from which the
embeddings should be extracted
batch_size: batch_size for inference
embedding_dir: path to directory to store embeddings file
device: compute device to perform operations
"""
all_embs, _, _, _ = speaker_model.batch_inference(manifest_file, batch_size=batch_size, device=device)
all_embs = np.asarray(all_embs)
all_embs = embedding_normalize(all_embs)
out_embeddings = {}
with open(manifest_file, 'r', encoding='utf-8') as manifest:
for i, line in enumerate(manifest.readlines()):
line = line.strip()
dic = json.loads(line)
uniq_name = '@'.join(dic['audio_filepath'].split('/')[-3:])
out_embeddings[uniq_name] = all_embs[i]
embedding_dir = os.path.join(embedding_dir, 'embeddings')
if not os.path.exists(embedding_dir):
os.makedirs(embedding_dir, exist_ok=True)
prefix = manifest_file.split('/')[-1].rsplit('.', 1)[-2]
name = os.path.join(embedding_dir, prefix)
embeddings_file = name + '_embeddings.pkl'
pkl.dump(out_embeddings, open(embeddings_file, 'wb'))
logging.info("Saved embedding files to {}".format(embedding_dir))
def main():
parser = ArgumentParser()
parser.add_argument(
"--manifest", type=str, required=True, help="Path to manifest file",
)
parser.add_argument(
"--model_path",
type=str,
default='titanet_large',
required=False,
help="path to .nemo speaker verification model file to extract embeddings, if not passed SpeakerNet-M model would be downloaded from NGC and used to extract embeddings",
)
parser.add_argument(
"--batch_size", type=int, default=1, required=False, help="batch size",
)
parser.add_argument(
"--embedding_dir",
type=str,
default='./',
required=False,
help="path to directory where embeddings need to stored default:'./'",
)
args = parser.parse_args()
torch.set_grad_enabled(False)
if args.model_path.endswith('.nemo'):
logging.info(f"Using local speaker model from {args.model_path}")
speaker_model = EncDecSpeakerLabelModel.restore_from(restore_path=args.model_path)
elif args.model_path.endswith('.ckpt'):
speaker_model = EncDecSpeakerLabelModel.load_from_checkpoint(checkpoint_path=args.model_path)
else:
speaker_model = EncDecSpeakerLabelModel.from_pretrained(model_name="titanet_large")
logging.info(f"using pretrained titanet_large speaker model from NGC")
device = 'cuda'
if not torch.cuda.is_available():
device = 'cpu'
logging.warning("Running model on CPU, for faster performance it is adviced to use atleast one NVIDIA GPUs")
get_embeddings(
speaker_model, args.manifest, batch_size=args.batch_size, embedding_dir=args.embedding_dir, device=device
)
if __name__ == '__main__':
main()
|