diffsptk

diffsptk is a differentiable version of SPTK based on the PyTorch framework.

Requirements

• Python 3.8+
• PyTorch 1.9.0+

Documentation

See this page for a reference manual.

Installation

The latest stable release can be installed through PyPI by running

pip install diffsptk

Alternatively,

git clone https://github.com/sp-nitech/diffsptk.git
pip install -e diffsptk

Examples

Mel-cepstral analysis

import diffsptk
import torch

x = torch.randn(100)

# Compute STFT of x.
stft = diffsptk.STFT(frame_length=12, frame_period=10, fft_length=16)
X = stft(x)

# Estimate 4-th order mel-cepstrum of x.
mcep = diffsptk.MelCepstralAnalysis(cep_order=4, fft_length=16, alpha=0.1, n_iter=1)
mc = mcep(X)

Mel-spectrogram extraction

import diffsptk
import torch

x = torch.randn(100)

# Compute STFT of x.
stft = diffsptk.STFT(frame_length=12, frame_period=10, fft_length=32)
X = stft(x)

# Apply 4 mel-filter banks to the STFT.
fbank = diffsptk.MelFilterBankAnalysis(n_channel=4, fft_length=32, sample_rate=8000)
Y = fbank(X)

Subband decomposition

import diffsptk
import torch

K = 4   # Number of subbands.
M = 40  # Order of filter.

x = torch.randn(100)

# Decompose x.
pqmf = diffsptk.PQMF(K, M)
decimate = diffsptk.Decimation(K)
y = decimate(pqmf(x), dim=-1)

# Reconstruct x.
interpolate = diffsptk.Interpolation(K)
ipqmf = diffsptk.IPQMF(K, M)
x_hat = ipqmf(interpolate(K * y, dim=-1))

# Compute error between two signals.
error = torch.abs(x_hat - x).sum()

Status

• acorr
• acr2csm
• aeq (torch.allclose)
• amgcep
• average (torch.mean)
• b2mc
• bcp
• bcut
• c2acr
• c2mpir
• c2ndps
• cdist
• clip (torch.clip)
• csm2acr
• dct
• decimate
• delay
• delta
• dequantize
• df2
• dfs
• dmp
• dtw
• dtw_merge
• entropy (torch.special.entr)
• excite
• extract
• fbank
• fd
• fdrw
• fft (torch.fft.fft)
• fft2 (torch.fft.fft2)
• fftcep
• fftr (torch.fft.rfft)
• fftr2 (torch.fft.rfft2)
• frame
• freqt
• glogsp
• gmm
• gmmp
• gnorm
• gpolezero
• grlogsp
• grpdelay
• gseries
• gspecgram
• gwave
• histogram (torch.histogram)
• huffman
• huffman_decode
• huffman_encode
• idct
• ifft (torch.fft.ifft)
• ifft2 (torch.fft.ifft2)
• ignorm
• imglsadf
• impulse
• imsvq
• interpolate
• ipqmf
• iulaw
• lar2par
• lbg
• levdur
• linear_intpl
• lpc
• lpc2c
• lpc2lsp
• lpc2par
• lpccheck
• lsp2lpc
• lspcheck
• lspdf
• ltcdf
• mc2b
• mcpf
• median (torch.median)
• merge
• mfcc
• mgc2mgc
• mgc2sp
• mgcep (mcep is now available)
• mglsadf
• mglsp2sp
• minmax
• mlpg
• mlsacheck
• mpir2c
• mseq
• msvq
• nan (torch.isnan)
• ndps2c
• norm0
• nrand (torch.randn)
• par2lar
• par2lpc
• pca
• pcas
• phase
• pitch
• pitch_mark
• poledf
• pqmf
• quantize
• ramp
• reverse
• rlevdur
• rmse
• root_pol
• sin
• smcep
• snr
• sopr
• spec
• step
• swab
• symmetrize
• train
• transpose
• ulaw
• vc
• vopr
• vstat (torch.var_mean)
• vsum (torch.sum)
• window
• x2x
• zcross
• zerodf

License

This software is released under the Apache License 2.0.