bpdp 0.1.0

Epoch extraction with bandpass filter and dynamic progamming

BPDP: A PyTorch implementation of our BPDP epoch extraction algorithm.

Description

This package implementation the BPDP epoch extraction algorithm using Pytorch.

Install

This package can be installed using pip.

pip install bpdp

Example Usage:

import bpdp
import torchaudio

x, sr = torchaudio.load("example.wav")
epochs = bpdp.bpdp(x, sr)

Documentation

This package only contains a few functions so we just write the doc here.

def bpdp.bpdp(x, sr, wl_0=0.05, wl_1=0.002, f_lo=50.0, f_hi=550.0, beam_size=5, filter="bp1", resampler='makima', _probe=None) -> Tuple[int, ...]:

The main entry point of our bpdp algorithm.

Parameters:

• x(torch.Tensor): The input signal of shape(L).
• sr(int): The sample rate in Hz.
• wl_0(float): The window length of the high pass filter in ms (Default=0.05).
• wl_1(float): The window length of the low pass filter in ms (Default=0.002).
• f_lo(float): The minimum fundamental frequency in Hz (Default=50.0).
• f_hi(float): The maximum fundatental frequency in Hz (Default=550.0).
• beam_size(int): The beam size used in dynamic programming (Default=5).
• filter(str): The filtering method. Should be one of:
  • 'bp1': Apply one bandpass filter (Default).
  • 'bp2': Apply a highpass filter and a lowpass filter sequentially.
• resampler(str): The resample method. Should be one of:
  • 'akima': Akima piecewise cubic Hermite interpolation.
  • 'makima': Modified Akima piecewise cubic Hermite interpolation (Default).
  • 'fseries': Inteplation using the Fourier series.
• _probe(Tuple[int, int] or None): Print the decision between sample n0 to n1. The input should be a 2-tuple containing n0 and n1 (Default=None).

Returns:

• p(Tuple[int, ...]): The indices of epoch in sample.

def bpdp.filter.bp1(sr, wl_0, wl_1) -> torch.Tensor:

Return a bandpass filter using K6 window.

Parameters:

• sr(int): The sample rate $f_s$ in Hz.
• wl_0(float): The window length $T_0$ in msec.
• wl_1(float): The window length $T_1$ in msec.

Return:

• h(torch.Tensor): A bandpass filter of size $L=\lfloor T_0 f_s + 1\rfloor$

def bpdp.filter.bp2(sr, wl_0, wl_1) -> Tuple[torch.Tensor, torch.Tensor]:

Return a highpass filter and a lowpass filter using K6 window.

Parameters:

• sr(int): The sample rate $f_s$ in Hz.
• wl_0(float): The window length $T_0$ in msec.
• wl_1(float): The window length $T_1$ in msec.

Return:

• h(torch.Tensor): A highpass filter of size $L=\lfloor T_0 f_s + 1\rfloor$
• g(torch.Tensor): A lowpass filter of size $L=\lfloor T_1 f_s + 1\rfloor$

def bpdp.resample.resample(x, m, method='makima'):

Resample the $N$-periodic signal $x$ to $M$-periodic signal $y$.

Parameters:

• x(torch.Tensor): The input vector $x$ of size $N$.
• m(int): The output period $M$.
• method(str): The interpolation method. Should be one of:
  • 'akima': Akima piecewise cubic Hermite interpolation.
  • 'makima': Modified Akima piecewise cubic Hermite interpolation (Default).
  • 'fseries': Inteplation using the Fourier series.

Returns

• y(torch.Tensor): The output vector $y$ of size $M$.

def bpdp.resample.resample_akima(x, m):

Resample the $N$-periodic signal $x$ to $M$-periodic signal $y$ using the Akima spline interpolation.

Parameters:

• x(torch.Tensor): The input vector $x$ of size $N$.
• m(int): The output period $M$.

Returns

• y(torch.Tensor): The output vector $y$ of size $M$.

def bpdp.resample.resample_makima(x, m):

Resample the $N$-periodic signal $x$ to $M$-periodic signal $y$ using the modified Akima spline interpolation.

Parameters:

• x(torch.Tensor): The input vector $x$ of size $N$.
• m(int): The output period $M$.

Returns

• y(torch.Tensor): The output vector $y$ of size $M$.

def bpdp.resample.resample_fseries(x, m):

Resample the $N$-periodic signal $x$ to $M$-periodic signal $y$ using the Fourier series.

Parameters:

• x(torch.Tensor): The input vector $x$ of size $N$.
• m(int): The output period $M$.

Returns

• y(torch.Tensor): The output vector $y$ of size $M$.

def bpdp.window.k6.k6(t, wl = 1.0, dtype=torch.float32):

Calculate Kawahara's 6-terms cosine series at time $t$. The window's length $L$ is equal to wl. $$ h(t) = w(t) = \sum_{k=0}^{5} a_k \cos(k 2 \pi \frac{t}{L}) $$ The frequency response at zero frequency $\hat h (0) = a_0 L.$ The -3db Point is about $1.289 / L$.

Parameters:

• t(torch.Tensor): The time $t$ for calculation, can be in any shape.
• wl(float): The window length.
• dtype(torch.dtype): The data type of the output.

Returns:

• h(torch.Tensor): The output tensor. Its shape is same as input tensor $t$