Spectral Features

Spectral features are widely used in machine and deep learning research. The code shows implementations of basic features for voice/speech analysis.

• SpectralEntropy
• SpectralCentroid
• SpectralSpread
• SpectralSkewness
• SpectralKurtosis
• SpectralRolloffPoint
• SpectralCrest
• (WIP) SpectralFlux
• SpectralSlope
• SpectralFlatness
• (WIP) Ma, Y., Nishihara, A. Efficient voice activity detection algorithm using long-term spectral flatness measure. J AUDIO SPEECH MUSIC PROC. 2013, 87 (2013).

References

The code highly refers to the Matlab tutorial, and uses Pytorch and some functions of Speechbrain to achieve.

Prepare an Environment

git clone https://github.com/BrownsugarZeer/SpectralFeatures.git
python -m venv venv
venv\Scripts\activate.bat
pip install -r requirements.txt

Running an Experiment

My file path of waveform is <path_to_matlab>\MATLAB\R2019b\toolbox\audio\samples\Counting-16-44p1-mono-15secs.wav and has been downsampled from 44100 Hz to 16000 Hz.

• Spectral Entropy

x, fs = torchaudio.load("Counting-16-44p1-mono-15secs_16000.wav")
compute_feat = SpectralEntropy(sample_rate=fs)
spectr_feat = compute_feat(x)
plot_feature(x, fs, spectr_feat)

---

• Spectral Centroid

x, fs = torchaudio.load("Counting-16-44p1-mono-15secs_16000.wav")
compute_feat = SpectralCentroid(sample_rate=fs)
spectr_feat = compute_feat(x)
plot_feature(x, fs, spectr_feat)

---

• Spectral Spread

x, fs = torchaudio.load("Counting-16-44p1-mono-15secs_16000.wav")
compute_feat = SpectralSpread(sample_rate=fs)
spectr_feat = compute_feat(x)
plot_feature(x, fs, spectr_feat)

---

• Spectral Skewness

x, fs = torchaudio.load("Counting-16-44p1-mono-15secs_16000.wav")
compute_feat = SpectralSkewness(sample_rate=fs)
spectr_feat = compute_feat(x)
plot_feature(x, fs, spectr_feat)

---

• Spectral Kurtosis

x, fs = torchaudio.load("Counting-16-44p1-mono-15secs_16000.wav")
compute_feat = SpectralKurtosis(sample_rate=fs)
spectr_feat = compute_feat(x)
plot_feature(x, fs, spectr_feat)

---

• Spectral Rolloff Point

x, fs = torchaudio.load("Counting-16-44p1-mono-15secs_16000.wav")
compute_feat = SpectralRolloffPoint(sample_rate=fs)
spectr_feat = compute_feat(x)
plot_feature(x, fs, spectr_feat)

---

• Spectral Crest

x, fs = torchaudio.load("Counting-16-44p1-mono-15secs_16000.wav")
compute_feat = SpectralCrest(sample_rate=fs)
spectr_feat = compute_feat(x)
plot_feature(x, fs, spectr_feat)

---

• Spectral Flux

x, fs = torchaudio.load("Counting-16-44p1-mono-15secs_16000.wav")
compute_feat = SpectralFlux(sample_rate=fs)
spectr_feat = compute_feat(x)
plot_feature(x, fs, spectr_feat)

---

• Spectral Slope

x, fs = torchaudio.load("Counting-16-44p1-mono-15secs_16000.wav")
compute_feat = SpectralSlope(sample_rate=fs)
spectr_feat = compute_feat(x)
plot_feature(x, fs, spectr_feat)

---

• Spectral Flatness

x, fs = torchaudio.load("Counting-16-44p1-mono-15secs_16000.wav")
compute_feat = SpectralFlatness(sample_rate=fs)
spectr_feat = compute_feat(x)
plot_feature(x, fs, spectr_feat)

---

• Long-Term Spectral Flatness

x, fs = torchaudio.load("Counting-16-44p1-mono-15secs_16000.wav")
compute_feat = LongTermSpectralFlatness(sample_rate=fs)
spectr_feat = compute_feat(x)
plot_feature(x, fs, spectr_feat)

Issue

The readability of the code will sometimes leads to lower performance..