diffsptk

diffsptk is a differentiable version of SPTK built on the PyTorch framework. It provides various speech signal processing modules as PyTorch layers, allowing users to integrate classic signal processing algorithms directly into neural network architectures and optimize them through backpropagation.

Requirements

• Python 3.10+
• PyTorch 2.3.1+

Documentation

• Reference Manual - Detailed API documentation and module specifications.
• Interactive Tutorial (Google Colab) - Hands-on examples to get started with diffsptk in your browser.
• Conference Paper - Technical background and implementation details available on the ISCA Archive.

Installation

The latest stable release can be installed via PyPI:

pip install diffsptk

Alternatively, the development version can be installed directly from the GitHub repository:

pip install git+https://github.com/sp-nitech/diffsptk.git@master

Examples

Running on a GPU

import diffsptk

stft_params = {"frame_length": 400, "frame_period": 80, "fft_length": 512}

# Read waveform.
x, sr = diffsptk.read("assets/data.wav", device="cuda")

# Compute spectrogram using a nn.Module class.
X1 = diffsptk.STFT(**stft_params, device="cuda")(x)

# Compute spectrogram using a functional method.
X2 = diffsptk.functional.stft(x, **stft_params)

print(X1.allclose(X2))

Mel-cepstral analysis and synthesis

import diffsptk

fl = 400     # Frame length.
fp = 80      # Frame period.
n_fft = 512  # FFT length.
M = 24       # Mel-cepstrum dimensions.

# Read waveform.
x, sr = diffsptk.read("assets/data.wav")

# Compute STFT amplitude of x.
stft = diffsptk.STFT(frame_length=fl, frame_period=fp, fft_length=n_fft)
X = stft(x)

# Estimate mel-cepstrum of x.
alpha = diffsptk.get_alpha(sr)
mcep = diffsptk.MelCepstralAnalysis(
    fft_length=n_fft,
    cep_order=M,
    alpha=alpha,
    n_iter=10,
)
mc = mcep(X)

# Reconstruct x.
mlsa = diffsptk.MLSA(filter_order=M, frame_period=fp, alpha=alpha, taylor_order=20)
x_hat = mlsa(mlsa(x, -mc), mc)

# Write reconstructed waveform.
diffsptk.write("reconst.wav", x_hat, sr)

# Compute error.
error = (x_hat - x).abs().sum()
print(error)

# Extract pitch of x.
pitch = diffsptk.Pitch(
    frame_period=fp,
    sample_rate=sr,
    f_min=80,
    f_max=180,
    voicing_threshold=0.4,
    out_format="pitch",
)
p = pitch(x)

# Generate excitation signal.
excite = diffsptk.ExcitationGeneration(frame_period=fp)
e = excite(p)
n = diffsptk.nrand(x.size(0) - 1)

# Synthesize waveform.
x_voiced = mlsa(e, mc)
x_unvoiced = mlsa(n, mc)

# Output analysis-synthesis result.
diffsptk.write("voiced.wav", x_voiced, sr)
diffsptk.write("unvoiced.wav", x_unvoiced, sr)

WORLD analysis and synthesis

import diffsptk

fp = 80       # Frame period.
n_fft = 1024  # FFT length.

# Read waveform.
x, sr = diffsptk.read("assets/data.wav")

# Extract F0 of x, or prepare well-estimated F0.
pitch = diffsptk.Pitch(
    frame_period=fp,
    sample_rate=sr,
    f_min=80,
    f_max=180,
    voicing_threshold=0.4,
    out_format="f0",
)
f0 = pitch(x)

# Extract aperiodicity of x by D4C.
ap = diffsptk.Aperiodicity(
    frame_period=fp,
    sample_rate=sr,
    fft_length=n_fft,
    algorithm="d4c",
    out_format="a",
)
A = ap(x, f0)

# Extract spectral envelope of x by CheapTrick.
pitch_spec = diffsptk.PitchAdaptiveSpectralAnalysis(
    frame_period=fp,
    sample_rate=sr,
    fft_length=n_fft,
    algorithm="cheap-trick",
    out_format="power",
)
S = pitch_spec(x, f0)

# Reconstruct x.
world_synth = diffsptk.WorldSynthesis(
    frame_period=fp,
    sample_rate=sr,
    fft_length=n_fft,
)
x_hat = world_synth(f0, A, S)

# Write reconstructed waveform.
diffsptk.write("reconst.wav", x_hat, sr)

# Compute error.
error = (x_hat - x).abs().sum()
print(error)

LPC analysis and synthesis

import diffsptk

fl = 400  # Frame length.
fp = 80   # Frame period.
M = 24    # LPC dimensions.

# Read waveform.
x, sr = diffsptk.read("assets/data.wav")

# Estimate LPC of x.
frame = diffsptk.Frame(frame_length=fl, frame_period=fp)
window = diffsptk.Window(in_length=fl)
lpc = diffsptk.LPC(frame_length=fl, lpc_order=M, eps=1e-5)
a = lpc(window(frame(x)))

# Convert to inverse filter coefficients.
norm0 = diffsptk.AllPoleToAllZeroDigitalFilterCoefficients(filter_order=M)
b = norm0(a)

# Reconstruct x.
zerodf = diffsptk.AllZeroDigitalFilter(filter_order=M, frame_period=fp)
poledf = diffsptk.AllPoleDigitalFilter(filter_order=M, frame_period=fp)
x_hat = poledf(zerodf(x, b), a)

# Write reconstructed waveform.
diffsptk.write("reconst.wav", x_hat, sr)

# Compute error.
error = (x_hat - x).abs().sum()
print(error)

PARCOR analysis and synthesis

import diffsptk

fl = 400  # Frame length.
fp = 80   # Frame period.
M = 24    # PARCOR dimensions.

# Read waveform.
x, sr = diffsptk.read("assets/data.wav")

# Estimate PARCOR of x.
frame = diffsptk.Frame(frame_length=fl, frame_period=fp)
window = diffsptk.Window(in_length=fl)
lpc = diffsptk.LPC(frame_length=fl, lpc_order=M, eps=1e-5)
lpc2par = diffsptk.LinearPredictiveCoefficientsToParcorCoefficients(lpc_order=M)
k = lpc2par(lpc(window(frame(x))))

# Reconstruct x.
iltcdf = diffsptk.AllZeroLatticeDigitalFilter(filter_order=M, frame_period=fp)
ltcdf = diffsptk.AllPoleLatticeDigitalFilter(filter_order=M, frame_period=fp)
x_hat = ltcdf(iltcdf(x, k), k)

# Write reconstructed waveform.
diffsptk.write("reconst.wav", x_hat, sr)

# Compute error.
error = (x_hat - x).abs().sum()
print(error)

Mel spectrogram analysis and synthesis

import diffsptk

fl = 400         # Frame length.
fp = 80          # Frame period.
n_fft = 512      # FFT length.
n_channel = 128  # Number of channels.

# Read waveform.
x, sr = diffsptk.read("assets/data.wav")

# Compute STFT amplitude of x.
stft = diffsptk.STFT(frame_length=fl, frame_period=fp, fft_length=n_fft)
X = stft(x)

# Extract log-mel spectrogram.
fbank = diffsptk.FBANK(
    fft_length=n_fft,
    n_channel=n_channel,
    sample_rate=sr,
)
Y = fbank(X)

# Reconstruct linear spectrogram.
ifbank = diffsptk.IFBANK(
    n_channel=n_channel,
    fft_length=n_fft,
    sample_rate=sr,
)
X_hat = ifbank(Y)

# Reconstruct x.
griffin = diffsptk.GriffinLim(
    frame_length=fl,
    frame_period=fp,
    fft_length=n_fft,
)
x_hat = griffin(X_hat, out_length=x.size(0))

# Write reconstructed waveform.
diffsptk.write("reconst.wav", x_hat, sr)

# Compute error.
error = (x_hat - x).abs().sum()
print(error)

Subband decomposition

import diffsptk

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

# Read waveform.
x, sr = diffsptk.read("assets/data.wav")

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

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

# Write reconstructed waveform.
diffsptk.write("reconst.wav", x_hat, sr)

# Compute error.
error = (x_hat - x).abs().sum()
print(error)

Gammatone filter bank analysis and synthesis

import diffsptk

# Read waveform.
x, sr = diffsptk.read("assets/data.wav")

# Decompose x.
gammatone = diffsptk.GammatoneFilterBankAnalysis(sr)
y = gammatone(x)

# Reconstruct x.
igammatone = diffsptk.GammatoneFilterBankSynthesis(sr)
x_hat = igammatone(y).reshape(-1)

# Write reconstructed waveform.
diffsptk.write("reconst.wav", x_hat, sr)

# Compute error.
error = (x_hat - x).abs().sum()
print(error)

Fractional octave band analysis and synthesis

import diffsptk

# Read waveform.
x, sr = diffsptk.read("assets/data.wav")

# Decompose x.
oband = diffsptk.FractionalOctaveBandAnalysis(sr)
y = oband(x)

# Reconstruct x.
x_hat = y.sum(1).reshape(-1)

# Write reconstructed waveform.
diffsptk.write("reconst.wav", x_hat, sr)

# Compute error.
error = (x_hat - x).abs().sum()
print(error)

Constant-Q transform

import diffsptk
import librosa  # This is to get sample audio.

fp = 128  # Frame period.
K = 252   # Number of CQ-bins.
B = 36    # Number of bins per octave.

# Read waveform.
x, sr = diffsptk.read(librosa.ex("trumpet"))

# Transform x.
cqt = diffsptk.CQT(fp, sr, n_bin=K, n_bin_per_octave=B)
c = cqt(x)

# Reconstruct x.
icqt = diffsptk.ICQT(fp, sr, n_bin=K, n_bin_per_octave=B)
x_hat = icqt(c, out_length=x.size(0))

# Write reconstructed waveform.
diffsptk.write("reconst.wav", x_hat, sr)

# Compute error.
error = (x_hat - x).abs().sum()
print(error)

Modified discrete cosine transform

import diffsptk

fl = 512  # Frame length.

# Read waveform.
x, sr = diffsptk.read("assets/data.wav")

# Transform x.
mdct = diffsptk.MDCT(fl)
c = mdct(x)

# Reconstruct x.
imdct = diffsptk.IMDCT(fl)
x_hat = imdct(c, out_length=x.size(0))

# Write reconstructed waveform.
diffsptk.write("reconst.wav", x_hat, sr)

# Compute error.
error = (x_hat - x).abs().sum()
print(error)

Vector quantization

import diffsptk

K = 2  # Codebook size.
M = 4  # Order of vector.

# Prepare input.
x = diffsptk.nrand(M)

# Quantize x.
vq = diffsptk.VectorQuantization(M, K)
x_hat, indices, commitment_loss = vq(x)

# Compute error.
error = (x_hat - x).abs().sum()
print(error)

License

This software is released under the Apache License 2.0.

Citation

@InProceedings{sp-nitech2023sptk,
  author = {Takenori Yoshimura and Takato Fujimoto and Keiichiro Oura and Keiichi Tokuda},
  title = {{SPTK4}: An open-source software toolkit for speech signal processing},
  booktitle = {12th ISCA Speech Synthesis Workshop (SSW 2023)},
  pages = {211--217},
  year = {2023},
}