Training, evaluation, and inference of neural pitch and periodicity estimators in PyTorch. Includes the original code for the paper "Cross-domain Neural Pitch and Periodicity Estimation".
If you want to perform pitch estimation using a pretrained FCNF0++ model, run
pip install penn
If you want to train or use your own models, run
pip install penn[train]
Perform inference using FCNF0++
import penn
# Load audio
audio, sample_rate = torchaudio.load('test/assets/gershwin.wav')
# Here we'll use a 10 millisecond hopsize
hopsize = .01
# Provide a sensible frequency range given your domain and model
fmin = 30.
fmax = 1000.
# Choose a gpu index to use for inference. Set to None to use cpu.
gpu = 0
# If you are using a gpu, pick a batch size that doesn't cause memory errors
# on your gpu
batch_size = 2048
# Select a checkpoint to use for inference. Selecting None will
# download and use FCNF0++ pretrained on MDB-stem-synth and PTDB
checkpoint = None
# Centers frames at hopsize / 2, 3 * hopsize / 2, 5 * hopsize / 2, ...
center = 'half-hop'
# (Optional) Linearly interpolate unvoiced regions below periodicity threshold
interp_unvoiced_at = .065
# Infer pitch and periodicity
pitch, periodicity = penn.from_audio(
audio,
sample_rate,
hopsize=hopsize,
fmin=fmin,
fmax=fmax,
checkpoint=checkpoint,
batch_size=batch_size,
center=center,
interp_unvoiced_at=interp_unvoiced_at,
gpu=gpu)
"""Perform pitch and periodicity estimation
Args:
audio: The audio to extract pitch and periodicity from
sample_rate: The audio sample rate
hopsize: The hopsize in seconds
fmin: The minimum allowable frequency in Hz
fmax: The maximum allowable frequency in Hz
checkpoint: The checkpoint file
batch_size: The number of frames per batch
center: Padding options. One of ['half-window', 'half-hop', 'zero'].
interp_unvoiced_at: Specifies voicing threshold for interpolation
gpu: The index of the gpu to run inference on
Returns:
pitch: torch.tensor(
shape=(1, int(samples // penn.seconds_to_sample(hopsize))))
periodicity: torch.tensor(
shape=(1, int(samples // penn.seconds_to_sample(hopsize))))
"""
"""Perform pitch and periodicity estimation from audio on disk
Args:
file: The audio file
hopsize: The hopsize in seconds
fmin: The minimum allowable frequency in Hz
fmax: The maximum allowable frequency in Hz
checkpoint: The checkpoint file
batch_size: The number of frames per batch
center: Padding options. One of ['half-window', 'half-hop', 'zero'].
interp_unvoiced_at: Specifies voicing threshold for interpolation
gpu: The index of the gpu to run inference on
Returns:
pitch: torch.tensor(shape=(1, int(samples // hopsize)))
periodicity: torch.tensor(shape=(1, int(samples // hopsize)))
"""
"""Perform pitch and periodicity estimation from audio on disk and save
Args:
file: The audio file
output_prefix: The file to save pitch and periodicity without extension
hopsize: The hopsize in seconds
fmin: The minimum allowable frequency in Hz
fmax: The maximum allowable frequency in Hz
checkpoint: The checkpoint file
batch_size: The number of frames per batch
center: Padding options. One of ['half-window', 'half-hop', 'zero'].
interp_unvoiced_at: Specifies voicing threshold for interpolation
gpu: The index of the gpu to run inference on
"""
"""Perform pitch and periodicity estimation from files on disk and save
Args:
files: The audio files
output_prefixes: Files to save pitch and periodicity without extension
hopsize: The hopsize in seconds
fmin: The minimum allowable frequency in Hz
fmax: The maximum allowable frequency in Hz
checkpoint: The checkpoint file
batch_size: The number of frames per batch
center: Padding options. One of ['half-window', 'half-hop', 'zero'].
interp_unvoiced_at: Specifies voicing threshold for interpolation
gpu: The index of the gpu to run inference on
"""
python -m penn
--audio_files AUDIO_FILES [AUDIO_FILES ...]
[-h]
[--config CONFIG]
[--output_prefixes OUTPUT_PREFIXES [OUTPUT_PREFIXES ...]]
[--hopsize HOPSIZE]
[--fmin FMIN]
[--fmax FMAX]
[--checkpoint CHECKPOINT]
[--batch_size BATCH_SIZE]
[--center {half-window,half-hop,zero}]
[--interp_unvoiced_at INTERP_UNVOICED_AT]
[--gpu GPU]
required arguments:
--audio_files AUDIO_FILES [AUDIO_FILES ...]
The audio files to process
optional arguments:
-h, --help
show this help message and exit
--config CONFIG
The configuration file. Defaults to using FCNF0++.
--output_prefixes OUTPUT_PREFIXES [OUTPUT_PREFIXES ...]
The files to save pitch and periodicity without extension.
Defaults to audio_files without extensions.
--hopsize HOPSIZE
The hopsize in seconds. Defaults to 0.01 seconds.
--fmin FMIN
The minimum frequency allowed in Hz. Defaults to 31.0 Hz.
--fmax FMAX
The maximum frequency allowed in Hz. Defaults to 1984.0 Hz.
--checkpoint CHECKPOINT
The model checkpoint file. Defaults to ./penn/assets/checkpoints/fcnf0++.pt.
--batch_size BATCH_SIZE
The number of frames per batch. Defaults to 2048.
--center {half-window,half-hop,zero}
Padding options
--interp_unvoiced_at INTERP_UNVOICED_AT
Specifies voicing threshold for interpolation. Defaults to 0.1625.
--gpu GPU
The index of the gpu to perform inference on. Defaults to CPU.
python -m penn.data.download
Downloads and uncompresses the mdb and ptdb datasets used for training.
python -m penn.data.preprocess --config <config>
Converts each dataset to a common format on disk ready for training. You can optionally pass a configuration file to override the default configuration.
python -m penn.partition
Generates train, valid, and test partitions for mdb and ptdb.
Partitioning is deterministic given the same random seed. You do not need to
run this step, as the original partitions are saved in
penn/assets/partitions.
python -m penn.train --config <config> --gpu <gpu>
Trains a model according to a given configuration on the mdb and ptdb
datasets.
You can monitor training via tensorboard.
tensorboard --logdir runs/ --port <port> --load_fast true
To use the torchutil notification system to receive notifications for long
jobs (download, preprocess, train, and evaluate), set the
PYTORCH_NOTIFICATION_URL environment variable to a supported webhook as
explained in the Apprise documentation.
python -m penn.evaluate \
--config <config> \
--checkpoint <checkpoint> \
--gpu <gpu>
Evaluate a model. <checkpoint> is the checkpoint file to evaluate and <gpu>
is the GPU index.
python -m penn.plot.density \
--config <config> \
--true_datasets <true_datasets> \
--inference_datasets <inference_datasets> \
--output_file <output_file> \
--checkpoint <checkpoint> \
--gpu <gpu>
Plot the data distribution and inferred distribution for a given dataset and save to a jpg file.
python -m penn.plot.logits \
--config <config> \
--audio_file <audio_file> \
--output_file <output_file> \
--checkpoint <checkpoint> \
--gpu <gpu>
Plot the pitch posteriorgram of an audio file and save to a jpg file.
python -m penn.plot.thresholds \
--names <names> \
--evaluations <evaluations> \
--output_file <output_file>
Plot the periodicity performance (voiced/unvoiced F1) over mdb and ptdb as a
function of the voiced/unvoiced threshold. names are the plot labels to give
each evaluation. evaluations are the names of the evaluations to plot.
M. Morrison, C. Hsieh, N. Pruyne, and B. Pardo, "Cross-domain Neural Pitch and Periodicity Estimation," Submitted to IEEE Transactions on Audio, Speech, and Language Processing, <TODO - month> 2023.
@inproceedings{morrison2023cross,
title={Cross-domain Neural Pitch and Periodicity Estimation},
author={Morrison, Max and Hsieh, Caedon and Pruyne, Nathan and Pardo, Bryan},
booktitle={Submitted to IEEE Transactions on Audio, Speech, and Language Processing},
month={TODO},
year={2023}
}