Torchcrepe

Warnings

• gotcha Torchcrepe's default Viterbi decoding differs from the original CREPE (TensorFlow) implementation. It uses Viterbi decoding on the softmax output instead of a weighted average, which helps prevent double/half frequency errors but changes the default pitch estimation approach.
  Fix: Be aware of this default behavior. For specific use cases, explore options in `torchcrepe.decode` if you need to replicate the original CREPE's decoding or implement custom post-processing.
• gotcha CREPE models were not trained on silent audio. This can lead to the model assigning high confidence to pitch bins even in silent regions. You may observe spurious pitch predictions in quiet sections.
  Fix: Utilize `torchcrepe.threshold.Silence` to manually set periodicity (confidence) to zero in silent regions, or apply custom silence detection and masking.
• gotcha The `batch_size` argument in `torchcrepe.predict` refers to internal batching over audio frames, not directly to processing multiple distinct audio files in a single call. Feeding multiple audio files of varying lengths in a batch for `predict` is not straightforward and might not offer the expected speed benefits due to padding overhead and other design choices.
  Fix: Process individual audio files separately or manage custom padding and batching strategies if you need to run multiple audio signals through the model concurrently. The library's `predict_from_files_to_files` functions are designed for convenience with multiple files, handling them sequentially.

Install

• pip install torchcrepe Install stable version

Imports

• torchcrepe

  import torchcrepe

• predict

  from torchcrepe import predict

  Main function for pitch prediction
• load.audio

  from torchcrepe.load import audio

  Utility to load audio files for processing

Quickstart

This quickstart demonstrates how to load an audio signal (using a mocked function for a self-contained example), set common parameters like hop length, frequency range, model capacity, and device, and then use `torchcrepe.predict` to estimate the pitch. It highlights the basic workflow for integrating torchcrepe into a PyTorch-based audio processing pipeline.

import torch
import torchcrepe
import numpy as np

# Mock torchcrepe.load.audio for a runnable example without external files
class MockLoadAudio:
    def audio(self, *args, **kwargs):
        # Generate a dummy 16kHz sine wave audio (1 second)
        sr = 16000
        duration = 1.0
        frequency = 440.0 # Hz
        t = np.linspace(0., duration, int(sr * duration), endpoint=False)
        audio_np = 0.5 * np.sin(2 * np.pi * frequency * t).astype(np.float32)
        return torch.from_numpy(audio_np).unsqueeze(0), sr # unsqueeze for batch dimension

torchcrepe.load = MockLoadAudio()

# Load dummy audio
audio, sr = torchcrepe.load.audio('dummy.wav', sr=16000)

# 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' if torch.cuda.is_available() else 'cpu'

# Pick a batch size that doesn't cause memory errors on your gpu
batch_size = 2048 # Note: Batching here refers to internal frame processing, not input audio files

# Compute pitch
pitch = torchcrepe.predict(
    audio,
    sr,
    hop_length,
    fmin,
    fmax,
    model,
    batch_size=batch_size,
    device=device,
    return_periodicity=False # Set to True to get a confidence score
)

print(f"Predicted pitch shape: {pitch.shape}")
if pitch.shape[-1] > 0:
    print(f"First few pitch values: {pitch[0, :5].tolist()}")