Pytorch implementation of the CREPE [1] pitch tracker. The original Tensorflow implementation can be found here. The provided model weights were obtained by converting the "tiny" and "full" models using MMdnn, an open-source model management framework.

Perform the system-dependent PyTorch install using the instructions found here.

pip install torchcrepe

import torchcrepe


# Load audio
audio, sr = torchcrepe.load.audio( ... )

# Here we'll use a 5 millisecond hop length
hop_length = int(sr / 200.)

# Provide a sensible frequency range for your domain (upper limit is 2006 Hz)
# This would be a reasonable range for speech
fmin = 50
fmax = 550

# Select a model capacity--one of "tiny" or "full"
model = 'tiny'

# Choose a device to use for inference
device = 'cuda:0'

# Pick a batch size that doesn't cause memory errors on your gpu
batch_size = 2048

# Compute pitch using first gpu
pitch = torchcrepe.predict(audio,
                           sr,
                           hop_length,
                           fmin,
                           fmax,
                           model,
                           batch_size=batch_size,
                           device=device)

A periodicity metric similar to the Crepe confidence score can also be extracted by passing return_periodicity=True to torchcrepe.predict.

By default, torchcrepe uses Viterbi decoding on the softmax of the network output. This is different than the original implementation, which uses a weighted average near the argmax of binary cross-entropy probabilities. The argmax operation can cause double/half frequency errors. These can be removed by penalizing large pitch jumps via Viterbi decoding. The decode submodule provides some options for decoding.

# Decode using viterbi decoding (default)
torchcrepe.predict(..., decoder=torchcrepe.decode.viterbi)

# Decode using weighted argmax (as in the original implementation)
torchcrepe.predict(..., decoder=torchcrepe.decode.weighted_argmax)

# Decode using argmax
torchcrepe.predict(..., decoder=torchcrepe.decode.argmax)

When periodicity is low, the pitch is less reliable. For some problems, it makes sense to mask these less reliable pitch values. However, the periodicity can be noisy and the pitch has quantization artifacts. torchcrepe provides submodules filter and threshold for this purpose. The filter and threshold parameters should be tuned to your data. For clean speech, a 10-20 millisecond window with a threshold of 0.21 has worked.

# We'll use a 15 millisecond window assuming a hop length of 5 milliseconds
win_length = 3

# Median filter noisy confidence value
periodicity = torchcrepe.filter.median(periodicity, win_length)

# Remove inharmonic regions
pitch = torchcrepe.threshold.At(.21)(pitch, periodicity)

# Optionally smooth pitch to remove quantization artifacts
pitch = torchcrepe.filter.mean(pitch, win_length)

For more fine-grained control over pitch thresholding, see torchcrepe.threshold.Hysteresis. This is especially useful for removing spurious voiced regions caused by noise in the periodicity values, but has more parameters and may require more manual tuning to your data.

CREPE was not trained on silent audio. Therefore, it sometimes assigns high confidence to pitch bins in silent regions. You can use torchcrepe.threshold.Silence to manually set the periodicity in silent regions to zero.

periodicity = torchcrepe.threshold.Silence(-60.)(periodicity,
                                                 audio,
                                                 sr,
                                                 hop_length)

batch = next(torchcrepe.preprocess(audio, sr, hop_length))
probabilities = torchcrepe.infer(batch)

As in Differentiable Digital Signal Processing [2], this uses the output of the fifth max-pooling layer as a pretrained pitch embedding

embeddings = torchcrepe.embed(audio, sr, hop_length)

torchcrepe defines the following functions convenient for predicting directly from audio files on disk. Each of these functions also takes a device argument that can be used for device placement (e.g., device='cuda:0').

torchcrepe.predict_from_file(audio_file, ...)
torchcrepe.predict_from_file_to_file(
    audio_file, output_pitch_file, output_periodicity_file, ...)
torchcrepe.predict_from_files_to_files(
    audio_files, output_pitch_files, output_periodicity_files, ...)

torchcrepe.embed_from_file(audio_file, ...)
torchcrepe.embed_from_file_to_file(audio_file, output_file, ...)
torchcrepe.embed_from_files_to_files(audio_files, output_files, ...)

usage: python -m torchcrepe
    [-h]
    --audio_files AUDIO_FILES [AUDIO_FILES ...]
    --output_files OUTPUT_FILES [OUTPUT_FILES ...]
    [--hop_length HOP_LENGTH]
    [--output_periodicity_files OUTPUT_PERIODICITY_FILES [OUTPUT_PERIODICITY_FILES ...]]
    [--embed]
    [--fmin FMIN]
    [--fmax FMAX]
    [--model MODEL]
    [--decoder DECODER]
    [--gpu GPU]
    [--no_pad]

optional arguments:
  -h, --help            show this help message and exit
  --audio_files AUDIO_FILES [AUDIO_FILES ...]
                        The audio file to process
  --output_files OUTPUT_FILES [OUTPUT_FILES ...]
                        The file to save pitch or embedding
  --hop_length HOP_LENGTH
                        The hop length of the analysis window
  --output_periodicity_files OUTPUT_PERIODICITY_FILES [OUTPUT_PERIODICITY_FILES ...]
                        The file to save periodicity
  --embed               Performs embedding instead of pitch prediction
  --fmin FMIN           The minimum frequency allowed
  --fmax FMAX           The maximum frequency allowed
  --model MODEL         The model capacity. One of "tiny" or "full"
  --decoder DECODER     The decoder to use. One of "argmax", "viterbi", or
                        "weighted_argmax"
  --gpu GPU             The gpu to perform inference on
  --no_pad              Whether to pad the audio

The module tests can be run as follows.

pip install pytest
pytest

[1] J. W. Kim, J. Salamon, P. Li, and J. P. Bello, “Crepe: A Convolutional Representation for Pitch Estimation,” in 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP).

[2] J. H. Engel, L. Hantrakul, C. Gu, and A. Roberts, “DDSP: Differentiable Digital Signal Processing,” in 2020 International Conference on Learning Representations (ICLR).