Signal transformation

Helps to transform signal data like speech into different representations using TensorFlow 2 API.

pip3 install signal-transformation

Mantains following types of output formats of the voice:

PCM
STFT
MAGNITUDE
MFCC
LOG MFCC

Voice

Transform wav files to needed format and then save them to tf_records

import pandas as pd
from signal_transformation.voice.tftr import wav_to_tf_records, SpectFormat    

metadata_path = '/path/to/parquet/file'
output_dir = '/output/directory'

metadata = pd.read_parquet(metadata_path)

wav_to_tf_records(
    metadata=metadata,
    output_dir=output_dir,
    spec_format=SpectFormat.MAGNITUDE
)

Or use as script

wav_to_tf_records --metadata_path="./vox1_metadata.gzip" --spect_format=3 --sample_rate=16000 --num_mfcc=13 --spect_shape="300, 200, 1" --num_shards=512 --output_dir="/dataset/tf_records"

Example of creating a metadata file of VoxCeleb 1

from signal_transformation import helpers   

helpers.voxceleb_files_to_metadata('files_dir', 'dev', 'vox1_metadata.gzip')

Or use as a script

vox_to_metadata --files_dir="/vox2/audio/dev/wav/" --part_name="dev" --output_file="/vox2/audio/dev/vox2_wav_metadata.gzip"

Create overlapping signals dataset

from signal_transformation import helpers


input_folder = '/path/to/wav/files'
output_folder = '/output/directory'

result = helpers.prepare_data(
    input_folder,
    output_folder,
    size=100,
    speakers_number=2,
    overlapping=True
)

Images

Imagenet dataset to TensorFlow records

imagenet_to_tf_records --mode="grayscale" --train_data_dirs "/datasets/images/imagenet/train_1/raw /datasets/images/imagenet/train_2/raw" --validation_data_dir /datasets/images/imagenet/validation --validation_labels_file /datasets/images/imagenet/ILSVRC2012_devkit_t12/data/ILSVRC2012_validation_ground_truth.txt --output_dir /datasets/images/imagenet/tf_records

Ilyushin/signal-transformation

Signal transformation

Voice

Images