Registry / ai-ml / torch-pitch-shift

torch-pitch-shift

library1.2.5pypypi✓ verified 25d ago

torch-pitch-shift is a Python library that enables rapid pitch-shifting of audio clips using PyTorch, with full CUDA support. It also provides utilities for calculating efficient pitch-shift targets, which is particularly useful for augmentation scenarios where speed is prioritized over precise pitch-shifts. The library is currently at version 1.2.5 and maintains an active development and release cadence.

ai-ml

pip install torch-pitch-shift

github ↗package ↗

INSTALL

IMPORT

SIG · TORCH-PITCH-SHIFT

torch-pitch-shift

ai-mlpythonv1.2.5

Install

65.8s avg

Import

5813ms

Disk

4736MB

Pass rate

4/ 10

Env Coverage4 / 10

glibc

3.9–3.13

musl

3.9–3.13

Install & Compatibility

Where this runs

tested against v1.2.5 · pip install

no network on importno background threads

Install × environment matrix

Each cell = how many times install + import succeeded across repeated harness runs. Partial = flaky.
glibc = Debian/Ubuntu slim · musl = Alpine Linux

musl

glibc

py 3.10

✕ build_error

✓ 74.8s

py 3.11

✕ build_error

✓ 69.7s

py 3.12

✕ build_error

✓ 61.7s

py 3.13

✕ build_error

✓ 57.1s

py 3.9

✕ build_error

✕ timeout

4736MB installed

● package 4736MB

Code

Verified usage

Verified import paths — ran on the pinned version, not inferred.

pitch_shift

✓ from torch_pitch_shift import pitch_shift

get_fast_shifts

✓ from torch_pitch_shift import get_fast_shifts

functional.pitch_shift

✓ from torch_pitch_shift import pitch_shift

✗ from torchaudio.functional import pitch_shift

The `torch-pitch-shift` library provides its own `pitch_shift` function, which has different arguments and internal implementation from `torchaudio.functional.pitch_shift`. Ensure you are importing from `torch_pitch_shift` if you intend to use this library's specific functionality.

This quickstart demonstrates how to generate a dummy audio waveform, apply a pitch shift of 3 semitones using the `pitch_shift` function, and print the resulting tensor shape. In a practical application, `waveform` would be loaded from an audio file (e.g., using `torchaudio.load`).

import torch from torch_pitch_shift import pitch_shift # Create a dummy mono audio waveform (batch_size, channels, samples) sample_rate = 44100 duration_seconds = 2 num_samples = sample_rate * duration_seconds waveform = torch.randn(1, 1, num_samples, dtype=torch.float32) # Define the pitch shift amount in semitones shift_semitones = 3.0 # Shift up by 3 semitones # Perform the pitch shift pitch_shifted_waveform = pitch_shift(waveform, shift_semitones, sample_rate) print(f"Original waveform shape: {waveform.shape}") print(f"Pitch-shifted waveform shape: {pitch_shifted_waveform.shape}") # In a real scenario, you would now save or play 'pitch_shifted_waveform'

Debug

Known issues

breakingOlder versions of `torch-pitch-shift` had specific compatibility requirements with `torchaudio` versions. For instance, v1.2.2 and v1.2.1 introduced backwards compatibility and explicit support for `torchaudio<=0.11.0` and `torchaudio v0.11` respectively. Users experiencing issues should check their `torchaudio` version and update `torch-pitch-shift` accordingly.

fix

Upgrade to `torch-pitch-shift>=1.2.2` and ensure `torchaudio` is within a compatible range, or explicitly manage `torchaudio` version based on `torch-pitch-shift` release notes.

affects: <1.2.2

gotchaFor improved audio quality and to reduce distortion during pitch-shifting, ensure you are using `torch-pitch-shift` version 1.2.0 or newer. This version introduced the `hop_length` argument, which significantly impacts output quality.

fix

Upgrade to `torch-pitch-shift>=1.2.0` and consider utilizing the `hop_length` argument for better results.

affects: <1.2.0

gotchaThe `get_fast_shifts` utility function may fail to compute valid pitch-shift ratios for certain sample rates or transpose ranges, raising an error. This typically occurs when no efficient fractional shifts can be found for the given parameters.

fix

Review the `sample_rate` and the `validate_fn` (transpose range) arguments passed to `get_fast_shifts`. Adjust the range or accept that not all desired shifts may be 'fast' for specific audio characteristics.

affects: All

gotchaThe library has been flagged for using dynamic code execution (e.g., `eval()`) by security analysis tools like Socket.dev. While not necessarily a vulnerability, this practice can pose security risks and may prevent the code from running in environments with strict security policies.

fix

Be aware of this characteristic if deploying in highly secure or constrained environments. Review the source code for specific `eval()` usages if concerns arise.

affects: All

gotchaWhile `torch-pitch-shift` is designed for speed with PyTorch/CUDA, performing pitch shifts on CPU can be significantly slower. Related libraries have also noted performance bottlenecks on CPU for similar operations, which might apply here.

fix

Prioritize GPU usage for performance-critical applications. If CPU usage is unavoidable, profile your code and consider batching operations to mitigate slowdowns.

affects: All

Errors

Common errors & fixes

ModuleNotFoundError: No module named 'torch_pitch_shift'

The `torch-pitch-shift` library is not installed in the current Python environment.

fix

Install the library using pip: `pip install torch-pitch-shift`

RuntimeError: Expected object of device type CUDA but got device type CPU for argument #1 'self'

There is a mismatch between the device of the input tensors (audio, shift_val) and the device on which the `torch_pitch_shift` operation is attempting to execute.

fix

Ensure all input tensors are explicitly moved to the same device (e.g., `'cuda'` or `'cpu'`) as the `torch_pitch_shift` operation, using `.to('cuda')` or `.to(device)`.

RuntimeError: Expected a floating point tensor, got a Long tensor for argument #2 'shift_val'

The `shift_val` argument, which specifies the pitch shift, was provided as an integer (Long) tensor instead of a floating-point tensor.

fix

Cast the `shift_val` tensor to a floating-point type, for example, by adding `.float()` or ensuring its initial creation uses float values: `torch.tensor([2.0])`.

RuntimeError: Given groups=1, weight of size [3, 1, 100], expected input[1, 1, 65536] to have 1 channels, but got 65536 channels instead

The input audio tensor has an incorrect shape for the operation. The library expects mono audio as `(batch_size, 1, time_samples)` but received a tensor (e.g., `(batch_size, time_samples)`) where the time dimension was implicitly interpreted as the channel dimension.

fix

Reshape the input audio tensor to explicitly include a channel dimension, typically `(batch_size, 1, time_samples)` for mono audio, using `tensor.unsqueeze(1)` if your `tensor` is `(batch_size, time_samples)`.

Upgrade

Version history

1.2.5latest on PyPI · released Sep 25, 2024

Audit

Dependencies

torchrequiredCore deep learning framework for tensor operations and GPU acceleration.

torchaudiorequiredAudio I/O and transformations, frequently updated with specific version compatibility requirements.

packagingrequiredUsed for version handling and compatibility checks within the library.

primepyrequiredA utility for prime number calculations, likely used in shift ratio computations.

Agent activity

15 hits · last 30 days

node

OpenAI (training)

Resources

githubhttps://github.com/KentoNishi/torch-pitch-shift ↗

packagehttps://pypi.org/project/torch-pitch-shift/ ↗