AudioHelper Module

Provides utilities for audio processing and manipulation.

class HMB.AudioHelper.AudiosHelper

Bases: object

AudiosHelper: Convenience methods for audio I/O and feature extraction.

This helper provides wrappers around librosa and spafe feature extractors as well as parselmouth-based voice feature extraction. Methods try to be thin wrappers and keep the original behavior; spafe-based methods accept a sample-rate (sr / fs) and forward it to the underlying library.

Notes

• Librosa is used for audio I/O and many feature extractions.

• Spafe is used for additional feature extractions not covered by librosa.

• You can install librosa and spafe via pip if not already installed: pip install librosa spafe praat-parselmouth

GetDuration(filePath)

Get the duration (in seconds) of an audio file.

Parameters:

filePath (str) – Path to the audio file.

Returns:

Duration in seconds as reported by librosa.

Return type:

float

References

• Librosa get_duration documentation: https://librosa.org/doc/latest/generated/librosa.get_duration.html

GetSegmentDuration(y, sr, roundTo=3)

Compute the duration of an audio signal array.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate of y (Hz).

• roundTo (int) – Number of decimals to round the result to (default 3).

Returns:

Rounded duration in seconds.

Return type:

float

References

• Librosa get_duration documentation: https://librosa.org/doc/latest/generated/librosa.get_duration.html

Load(filePath, conversionType=True, offset=0, segmentDuration=1, isReversed=False)

Load an audio file segment using librosa.

Parameters:

• filePath (str) – Path to the audio file.

• conversionType (bool) – If True, force mono. Matches librosa’s mono parameter.

• offset (float) – Start reading after this time (in seconds).

• segmentDuration (float) – Duration to load (in seconds). If 0 or None, librosa loads full file.

• isReversed (bool) – If True, reverse the returned signal (y[::-1]).

Returns:

(y, sr) where y is a 1-D numpy array and sr is the sampling rate (int).

• numpy.ndarray: Audio time series.

• int: Sampling rate of y.

Return type:

tuple

References

• Librosa load documentation: https://librosa.org/doc/latest/generated/librosa.load.html

GetSTFT(y)

Compute the short-time Fourier transform (STFT) of a signal.

Parameters:

y (numpy.ndarray) – 1-D audio time series.

Returns:

Complex-valued STFT matrix as returned by librosa.stft.

Return type:

numpy.ndarray

References

• Librosa STFT: https://librosa.org/doc/latest/generated/librosa.stft.html

• General STFT description: short-time Fourier transform literature.

GetAbsoluteSTFT(y)

Compute the magnitude (absolute value) spectrogram from the STFT.

Parameters:

y (numpy.ndarray) – 1-D audio time series.

Returns:

Magnitude spectrogram (non-negative floats).

Return type:

numpy.ndarray

References

• STFT magnitude and spectrogram conversions (librosa).

GetHarmonicEffect(y)

Extract the harmonic component of a signal using librosa.effects.harmonic.

Parameters:

y (numpy.ndarray) – 1-D audio time series.

Returns:

Harmonic component of the signal.

Return type:

numpy.ndarray

References

• Librosa harmonic/percussive separation: https://librosa.org/doc/latest/generated/librosa.effects.harmonic.html

GetPercussiveEffect(y)

Extract the percussive component of a signal using librosa.effects.percussive.

Parameters:

y (numpy.ndarray) – 1-D audio time series.

Returns:

Percussive component of the signal.

Return type:

numpy.ndarray

References

• Librosa harmonic/percussive separation: https://librosa.org/doc/latest/generated/librosa.effects.percussive.html

GetSlaneyMFCC(y, sr=22050, nMFCC=None)

Compute Slaney-style MFCCs using librosa.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate of y (Hz). Default is 22050.

• nMFCC (int, optional) – Number of MFCC coefficients to return. If None uses librosa default.

Returns:

MFCC matrix (n_mfcc x frames).

Return type:

numpy.ndarray

References

• MFCC background and Slaney implementation: Milner & Shao (2006) and Librosa MFCC docs.

• Librosa mfcc: https://librosa.org/doc/latest/generated/librosa.feature.mfcc.html

• Milner, B., & Shao, X. (2006). Clean speech reconstruction from MFCC vectors and fundamental frequency using an integrated front-end.

GetMeanSlaneyMFCC(y, sr, nMFCC=None)

Compute the mean (per-coefficient) Slaney MFCC over time.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate of y (Hz).

• nMFCC (int, optional) – Number of MFCC coefficients to average.

Returns:

1-D array of length n_mfcc containing the mean over frames.

Return type:

numpy.ndarray

References

• Librosa MFCC and averaging practices: https://librosa.org/doc/latest/generated/librosa.feature.mfcc.html

GetHtkMFCC(y, sr, nMFCC=None)

Compute HTK-style MFCCs using librosa with dct_type=3.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate of y (Hz).

• nMFCC (int, optional) – Number of MFCC coefficients to return.

Returns:

MFCC matrix (n_mfcc x frames).

Return type:

numpy.ndarray

References

• HTK MFCC conventions (DCT type 3) and Librosa MFCC settings.

• Librosa mfcc: https://librosa.org/doc/latest/generated/librosa.feature.mfcc.html

GetMeanHtkMFCC(y, sr, nMFCC=None)

Return mean HTK MFCC coefficients across time.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate of y (Hz).

• nMFCC (int, optional) – Number of MFCC coefficients to average.

Returns:

1-D array with mean HTK MFCC coefficients.

Return type:

numpy.ndarray

References

• HTK MFCC conventions and Librosa documentation.

GetMeanChroma(y, sr)

Compute mean chroma_stft features across time.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate of y (Hz).

Returns:

1-D array with 12 chroma mean values.

Return type:

numpy.ndarray

References

• Chroma feature analysis: Ellis (2007). See Librosa chroma_stft docs.

• Librosa chroma_stft: https://librosa.org/doc/latest/generated/librosa.feature.chroma_stft.html

GetMeanChromaSTFT(y, sr)

Compute mean chroma using the magnitude STFT as input.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate.

Returns:

Mean chroma vector.

Return type:

numpy.ndarray

References

• Chroma via STFT: Librosa documentation and chroma literature.

• Librosa chroma_stft: https://librosa.org/doc/latest/generated/librosa.feature.chroma_stft.html

GetMeanChromaCqt(y, sr)

Compute mean chroma from the constant-Q transform.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate.

Returns:

Mean chroma vector from CQT.

Return type:

numpy.ndarray

References

• Chroma CQT implementation: Librosa chroma_cqt docs.

• Muller, M. & Ewert, S. (2011) Chroma Toolbox reference.

GetMeanChromaCens(y, sr)

Compute chroma CENS mean across frames.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate.

Returns:

Mean chroma CENS vector.

Return type:

numpy.ndarray

References

• Chroma CENS method: Chroma Toolbox and related publications.

GetMeanMelSpectrogram(y, sr)

Compute mean Mel spectrogram across time.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate.

Returns:

Mean Mel spectrogram vector across frames.

Return type:

numpy.ndarray

References

• Librosa melspectrogram: https://librosa.org/doc/latest/generated/librosa.feature.melspectrogram.html

GetMeanSpectralContrast(y, sr)

Compute mean spectral contrast across time.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate.

Returns:

Mean spectral contrast vector across frames.

Return type:

numpy.ndarray

References

• Spectral contrast: Jiang et al. (2002). See Librosa spectral_contrast docs.

• Librosa spectral_contrast: https://librosa.org/doc/latest/generated/librosa.feature.spectral_contrast.html

GetMeanHarmonicTonnetz(y, sr)

Compute mean Tonnetz on the harmonic component.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate.

Returns:

Mean Tonnetz vector computed on harmonic component.

Return type:

numpy.ndarray

References

• Tonnetz features: Harte et al. (2006). See Librosa tonnetz docs.

• Librosa tonnetz: https://librosa.org/doc/latest/generated/librosa.feature.tonnetz.html

GetMeanTonnetz(y, sr)

Compute mean Tonnetz features across time.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate.

Returns:

Mean Tonnetz vector across frames.

Return type:

numpy.ndarray

References

• Tonnetz: Harte et al. (2006) and Librosa documentation.

GetMeanRMS(y, sr)

Compute mean root-mean-square energy across frames.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate.

Returns:

Mean RMS energy per frame.

Return type:

numpy.ndarray

References

• RMS energy and librosa RMS: https://librosa.org/doc/latest/generated/librosa.feature.rms.html

GetMeanSpectralCentroid(y, sr)

Compute mean spectral centroid across frames.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate.

Returns:

Mean spectral centroid per frame.

Return type:

numpy.ndarray

References

• Spectral centroid and bandwidth: Klapuri & Davy (2007). See Librosa docs.

• Librosa spectral_centroid: https://librosa.org/doc/latest/generated/librosa.feature.spectral_centroid.html

GetMeanSpectralBandwidth(y, sr)

Compute mean spectral bandwidth across frames.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate.

Returns:

Mean spectral bandwidth per frame.

Return type:

numpy.ndarray

References

• Spectral bandwidth discussion and Librosa docs.

• Librosa spectral_bandwidth: https://librosa.org/doc/latest/generated/librosa.feature.spectral_bandwidth.html

GetMeanSpectralRolloff(y, sr)

Compute mean spectral rolloff across frames.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate.

Returns:

Mean spectral rolloff per frame.

Return type:

numpy.ndarray

References

• Spectral rolloff: Librosa documentation.

• Librosa spectral_rolloff: https://librosa.org/doc/latest/generated/librosa.feature.spectral_rolloff.html

GetMeanSpectralFlatness(y, sr)

Compute mean spectral flatness across frames.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate.

Returns:

Mean spectral flatness per frame.

Return type:

numpy.ndarray

References

• Spectral flatness literature and Librosa docs.

• Dubnov et al. (2004) on spectral flatness.

GetMeanZCR(y, sr)

Compute mean zero-crossing rate across frames.

Parameters:

• y (numpy.ndarray) – 1-D audio time series.

• sr (int) – Sampling rate.

Returns:

Mean zero-crossing rate per frame.

Return type:

numpy.ndarray

References

• Zero-crossing Rate features and Librosa docs.

• Librosa zero_crossing_rate: https://librosa.org/doc/latest/generated/librosa.feature.zero_crossing_rate.html

GenerateScaledMelSpectrogram(y, sr, hopLength=512, nFFT=2048, numMels=128)

Create a log-scaled Mel spectrogram (dB) suitable for visualization or model input.

Parameters:

• y (numpy.ndarray) – Audio time series.

• sr (int) – Sampling rate.

• hopLength (int) – Hop length for STFT.

• nFFT (int) – FFT size.

• numMels (int) – Number of Mel bands to generate.

Returns:

Mel spectrogram in decibels (shape: n_mels x frames).

Return type:

numpy.ndarray

References

• Mel spectrogram and power_to_db: Librosa documentation.

• Librosa melspectrogram: https://librosa.org/doc/latest/generated/librosa.feature.melspectrogram.html

• Librosa power_to_db: https://librosa.org/doc/latest/generated/librosa.power_to_db.html

GenerateSTFT(y, sr, hopLength=512, nFFT=2048)

Compute a log-amplitude STFT spectrogram.

Parameters:

• y (numpy.ndarray) – Audio time series.

• sr (int) – Sampling rate.

• hopLength (int) – Hop length for STFT.

• nFFT (int) – FFT size.

Returns:

Log-amplitude STFT spectrogram.

Return type:

numpy.ndarray

References

• STFT and amplitude-to-db: Librosa documentation.

• Librosa amplitude_to_db: https://librosa.org/doc/latest/generated/librosa.amplitude_to_db.html

GetBFCC(y, sr=16000)

Compute BFCC (bark-frequency cepstral coefficients) using spafe.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate (fs) forwarded to spafe.bfcc.

Returns:

BFCC feature matrix.

Return type:

numpy.ndarray

References

• BFCC paper: https://asmp-eurasipjournals.springeropen.com/track/pdf/10.1186/s13636-017-0100-x

GetGFCC(y, sr=16000)

Compute GFCC (gammatone-frequency cepstral coefficients) using spafe.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

GFCC feature matrix.

Return type:

numpy.ndarray

References

• GFCC reference: https://www.researchgate.net/publication/309149564_Robust_Speaker_Verification_Using_GFCC_Based_i-Vectors

GetLFCC(y, sr=16000)

Compute LFCC (linear-frequency cepstral coefficients) using spafe.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

LFCC feature matrix.

Return type:

numpy.ndarray

References

• LFCC reference: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.63.8029&rep=rep1&type=pdf

GetLPC(y, sr=16000)

Compute LPC coefficients using spafe.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

LPC coefficient matrix.

Return type:

numpy.ndarray

References

• Linear predictive coding (LPC) literature and implementations.

GetLPCC(y, sr=16000)

Compute LPCC coefficients using spafe.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

LPCC coefficient matrix.

Return type:

numpy.ndarray

References

• LPCC and LP-based cepstral literature.

GetMFCC(y, sr=16000)

Compute MFCC using spafe (not librosa’s MFCC).

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

MFCC feature matrix from spafe.

Return type:

numpy.ndarray

References

• MFCC literature and implementations. See Milner & Shao (2006) and Librosa docs.

GetIMFCC(y, sr=16000)

Compute IMFCC (inverse MFCC) using spafe.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

IMFCC feature matrix.

Return type:

numpy.ndarray

References

• IMFCC reference literature and spafe implementation.

GetMSRCC(y, sr=16000)

Compute MSRCC features using spafe.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

MSRCC feature matrix.

Return type:

numpy.ndarray

References

• MSRCC reference: http://www.apsipa.org/proceedings/2018/pdfs/0001945.pdf

GetNGCC(y, sr=16000)

Compute NGCC features using spafe.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

NGCC feature matrix.

Return type:

numpy.ndarray

References

• NGCC references and related publications.

GetPNCC(y, sr=16000)

Compute PNCC features using spafe.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

PNCC feature matrix.

Return type:

numpy.ndarray

References

• PNCC implementation reference: https://github.com/supikiti/PNCC/blob/master/pncc.py

GetPSRCC(y, sr=16000)

Compute PSRCC features using spafe.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

PSRCC feature matrix.

Return type:

numpy.ndarray

References

• PSRCC reference: http://www.apsipa.org/proceedings/2018/pdfs/0001945.pdf

GetPLP(y, sr)

Compute PLP features using spafe.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

PLP feature matrix.

Return type:

numpy.ndarray

References

• PLP literature and spafe usage notes.

GetRPLP(y, sr)

Compute RPLP features using spafe.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

RPLP feature matrix.

Return type:

numpy.ndarray

References

• RPLP references and spafe implementation.

GetMeanBFCC(y, sr=16000)

Return mean BFCC coefficients across time (per-coefficient mean).

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

Mean BFCC vector.

Return type:

numpy.ndarray

References

• BFCC paper: https://asmp-eurasipjournals.springeropen.com/track/pdf/10.1186/s13636-017-0100-x

GetMeanGFCC(y, sr=16000)

Return mean GFCC coefficients across time.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

Mean GFCC vector.

Return type:

numpy.ndarray

References

• GFCC reference: https://www.researchgate.net/publication/309149564_Robust_Speaker_Verification_Using_GFCC_Based_i-Vectors

GetMeanLFCC(y, sr=16000)

Return mean LFCC coefficients across time.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

Mean LFCC vector.

Return type:

numpy.ndarray

References

• LFCC reference: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.63.8029&rep=rep1&type=pdf

GetMeanLPC(y, sr=16000)

Return mean LPC coefficients across time.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

Mean LPC vector.

Return type:

numpy.ndarray

References

• LPC refernece: https://superkogito.github.io/spafe/features/lpc.html

GetMeanLPCC(y, sr=16000)

Return mean LPCC coefficients across time.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

Mean LPCC vector.

Return type:

numpy.ndarray

References

• LPCC literature on cepstral analysis.

GetMeanMFCC(y, sr=16000)

Return mean MFCC coefficients (from spafe-mfcc) across time.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

Mean MFCC vector.

Return type:

numpy.ndarray

References

• MFCC reference: https://spafe.readthedocs.io/en/latest/features/mfcc.html

• Milner & Shao (2006) reference for MFCC reconstruction.

GetMeanIMFCC(y, sr=16000)

Return mean IMFCC coefficients across time.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

Mean IMFCC vector.

Return type:

numpy.ndarray

References

• IMFCC literature and spafe docs.

GetMeanMSRCC(y, sr=16000)

Return mean MSRCC coefficients across time.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

Mean MSRCC vector.

Return type:

numpy.ndarray

References

• MSRCC reference: http://www.apsipa.org/proceedings/2018/pdfs/0001945.pdf

GetMeanNGCC(y, sr=16000)

Return mean NGCC coefficients across time.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

Mean NGCC vector.

Return type:

numpy.ndarray

References

• NGCC references in the literature.

GetMeanPNCC(y, sr=16000)

Return mean PNCC coefficients across time.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

Mean PNCC vector.

Return type:

numpy.ndarray

References

• PNCC implementation: https://github.com/supikiti/PNCC/blob/master/pncc.py

GetMeanPSRCC(y, sr=16000)

Return mean PSRCC coefficients across time.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

Mean PSRCC vector.

Return type:

numpy.ndarray

References

• PSRCC reference: http://www.apsipa.org/proceedings/2018/pdfs/0001945.pdf

GetMeanPLP(y, sr)

Return mean PLP coefficients across time.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

Mean PLP vector.

Return type:

numpy.ndarray

References

• PLP literature.

GetMeanRPLP(y, sr)

Return mean RPLP coefficients across time.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

Mean RPLP vector.

Return type:

numpy.ndarray

References

• RPLP and rasta-PLP references.

GetAllMeanAudioFeatures(y, sr=16000)

Compute all mean audio features available in this class.

Parameters:

• y (numpy.ndarray) – Signal.

• sr (int) – Sample rate.

Returns:

Dictionary with mean feature vectors for all available features.

Return type:

dict

ExtractAudioFeaturesViaParselmouth(voiceSample, f0Min, f0Max, unit)

Extract voice features (pitch, jitter, shimmer, HNR, MFCC means) using parselmouth (Praat bindings).

Parameters:

• voiceSample (str or numpy.ndarray) – Path to audio file or a compatible array accepted by parselmouth.Sound.

• f0Min (float) – Minimum pitch (Hz) for pitch extraction.

• f0Max (float) – Maximum pitch (Hz) for pitch extraction.

• unit (str) – Unit used by Praat for mean/std extraction (e.g., ‘Hertz’).

Returns:

Dictionary with keys for f0 mean/std, hnr, jitter/shimmer measures, duration and mfccMean coefficients.

Return type:

dict

References

• Parselmouth (Praat bindings) documentation: https://parselmouth.readthedocs.io/en/stable/

GetMelImage(y, sr=16000, numMels=128, hopLength=512, nFFT=2048, outFrames=128, axisLast=True)

Generate a Mel spectrogram image (2D numpy array) from audio time series.

Parameters:

• y (numpy.ndarray) – Audio time series.

• sr (int) – Sampling rate.

• numMels (int) – Number of Mel bands.

• hopLength (int) – Hop length for STFT.

• nFFT (int) – FFT size.

• outFrames (int) – Desired number of output frames (time dimension).

Returns:

Mel spectrogram image with shape (numMels, outFrames, 3) if axisLast is True, else (3, numMels, outFrames).

Return type:

numpy.ndarray