audiofeat

A comprehensive PyTorch-based audio feature extraction library for speech research, music analysis, and audio ML pipelines. Extract 140+ features across temporal, spectral, cepstral, pitch, voice-quality, and rhythm domains — from a single pip install.

import audiofeat

features = audiofeat.extract_features_from_file("recording.wav")
print(features["f0_mean_hz"], features["rms_mean"], features["mfcc_0_mean"])

Why audiofeat?

• One library, all features. Temporal, spectral, cepstral, pitch, voice quality, rhythm, formants, and tonal features in a single package.
• PyTorch-first. Every feature returns a torch.Tensor. Plug directly into your training loop — no numpy-to-tensor conversion needed.
• Librosa-grade accuracy. Primary paths delegate to librosa when available for bit-exact parity; pure-PyTorch fallbacks when it's not installed.
• Beginner to production. Use individual functions for exploration, or the built-in CLI and batch extraction for production pipelines.
• Validated. Built-in Praat comparison tooling and a gold-standard scorecard for reproducible research.

Features

Temporal Features

Feature	Function	Description
RMS	rms()	Root-mean-square amplitude per frame
Short-Time Energy	short_time_energy()	Sum of squared signal values in each frame
Zero-Crossing Rate	zero_crossing_rate()	Rate at which the signal changes sign
Zero-Crossing Count	zero_crossing_count()	Number of zero-crossings per frame
Loudness	loudness()	Perceptual loudness estimation
Log Attack Time	log_attack_time()	MPEG-7 style attack time (10%–90% rise)
Decay Time	decay_time()	Time for envelope to decay from peak
Temporal Centroid	temporal_centroid()	Center of gravity of the amplitude envelope
Amplitude Modulation	amplitude_modulation_depth()	Depth of amplitude modulation over a sliding window
Entropy of Energy	entropy_of_energy()	Abrupt changes in energy within a frame
Teager Energy	teager_energy_operator()	Teager-Kaiser energy for amplitude/frequency tracking
Breath Group Duration	breath_group_duration()	Estimated duration of breath groups
Speech Rate	speech_rate()	Syllables per second estimation
Tristimulus	tristimulus()	T1/T2/T3 timbre ratios from harmonic amplitudes

Spectral Features

Feature	Function	Description
Spectral Centroid	spectral_centroid()	Center of mass of the spectrum
Spectral Rolloff	spectral_rolloff()	Frequency below which X% of energy is concentrated
Spectral Flux	spectral_flux()	Rate of change of the power spectrum
Spectral Flatness	spectral_flatness()	How noise-like a sound is (Wiener entropy)
Spectral Entropy	spectral_entropy()	Randomness of the spectral distribution
Spectral Bandwidth	spectral_bandwidth()	Spread of the spectrum around the centroid
Spectral Spread	spectral_spread()	Standard deviation of the spectral distribution
Spectral Slope	spectral_slope()	Linear regression slope fitted to the spectrum
Spectral Skewness	spectral_skewness()	Asymmetry of the spectral distribution
Spectral Crest Factor	spectral_crest_factor()	Peak-to-average ratio (peakiness)
Spectral Contrast	spectral_contrast()	Peak-valley amplitude difference across sub-bands
Spectral Deviation	spectral_deviation()	Jaggedness of the spectral envelope
Spectral Sharpness	spectral_sharpness()	Perceived sharpness (Zwicker model)
Spectral Roughness	spectral_roughness()	Sensory dissonance measure
Spectral Tonality	spectral_tonality()	Tonal vs. noise-like character
Spectral Irregularity	spectral_irregularity()	Irregularity of the spectral envelope
Low-High Energy Ratio	low_high_energy_ratio()	Energy below 1 kHz vs. above 3 kHz
HNR	harmonic_to_noise_ratio()	Harmonic-to-noise ratio
Harmonic Richness	harmonic_richness_factor()	Richness of harmonic content
Inharmonicity	inharmonicity_index()	Inharmonicity of the spectrum
Phase Coherence	phase_coherence()	Phase coherence across frequency bins
Sibilant Peak	sibilant_spectral_peak_frequency()	Peak frequency in the sibilant region

Spectrograms & Transforms

Feature	Function	Description
Linear Spectrogram	linear_spectrogram()	STFT magnitude spectrogram
Mel Spectrogram	mel_spectrogram()	Mel-scaled frequency spectrogram
Log Mel Spectrogram	log_mel_spectrogram()	Log-scaled Mel spectrogram
CQT Spectrogram	cqt_spectrogram()	Constant-Q transform (log-frequency bins)
MFCCs	mfcc()	Mel-Frequency Cepstral Coefficients
Chroma	chroma()	12-bin pitch class intensity (chromagram)
Tonnetz	tonnetz()	6D tonal centroid features

Formant Analysis

Feature	Function	Description
Formant Frequencies	formant_frequencies()	Extract F1, F2, F3, ... via Burg LPC
Formant Contours	formant_contours()	Time-varying formant trajectories
Formant Bandwidths	formant_bandwidths()	Bandwidth of each formant
Formant Dispersion	formant_dispersion()	Average spacing between formants

Linear Prediction

Feature	Function	Description
LPC	lpc_coefficients()	Linear Prediction Coefficients (Burg method)
LSP	lsp_coefficients()	Line Spectral Pairs from LPC

Cepstral Features

Feature	Function	Description
LPCC	lpcc()	Linear Predictive Cepstral Coefficients
GTCC	gtcc()	Gammatone Cepstral Coefficients
GFCC	gfcc()	Gammatone Frequency Cepstral Coefficients
ERB Cepstral	erb_cepstral_coefficients()	ERB-scale cepstral coefficients
Delta	delta()	First-order derivative of a feature contour
Delta-Delta	delta_delta()	Second-order derivative (acceleration)

Pitch Features

Feature	Function	Description
F0 (Autocorrelation)	fundamental_frequency_autocorr()	F0 via autocorrelation
F0 (YIN)	fundamental_frequency_yin()	F0 via YIN algorithm
F0 (pYIN)	fundamental_frequency_pyin()	Probabilistic YIN (requires librosa)
F0 (Praat)	fundamental_frequency_praat()	Exact Praat parity (requires parselmouth)
Pitch Strength	pitch_strength()	Strength of periodicity
Semitone Std Dev	semitone_sd()	F0 variation in semitones

Voice Quality Features

Feature	Function	Description
Jitter	jitter()	Cycle-to-cycle F0 variation
Jitter (local)	jitter_local()	Average absolute period difference (%)
Jitter (PPQ5)	jitter_ppq5()	Five-point Period Perturbation Quotient
Jitter (DDP)	jitter_ddp()	Difference of Differences of Periods
Shimmer	shimmer()	Cycle-to-cycle amplitude variation
Shimmer (local)	shimmer_local()	Local shimmer (%)
Shimmer (dB)	shimmer_local_db()	Shimmer in decibels
Shimmer (APQ3)	shimmer_apq3()	Three-point Amplitude Perturbation Quotient
Shimmer (DDA)	shimmer_dda()	Difference of Differences of Amplitudes
CPP	cepstral_peak_prominence()	Cepstral Peak Prominence for dysphonia detection
Alpha Ratio	alpha_ratio()	Energy ratio: 50–1000 Hz vs 1–5 kHz
Hammarberg Index	hammarberg_index()	Max energy ratio: 0–2 kHz vs 2–5 kHz
Harmonic Differences	harmonic_differences()	H1-H2, H1-A3, and other harmonic ratios
SHR	subharmonic_to_harmonic_ratio()	Subharmonic-to-harmonic power ratio
NAQ	normalized_amplitude_quotient()	Normalized Amplitude Quotient
Closed Quotient	closed_quotient()	Closed phase ratio from EGG
Soft Phonation Index	soft_phonation_index()	Low/high band energy ratio
GNE	glottal_to_noise_excitation()	Glottal-to-Noise Excitation ratio
MFDR	maximum_flow_declination_rate()	Maximum Flow Declination Rate
Vocal Fry Index	vocal_fry_index()	Ratio of fry frames to voiced frames
VOT	voice_onset_time()	Voice Onset Time estimation
Vocal Tract Length	vocal_tract_length()	Estimated from F1 and F2
Nasality Index	nasality_index()	Nasal vs. oral microphone energy

Rhythm Features

Feature	Function	Description
Tempo	tempo()	BPM estimation from onset autocorrelation
Beat Tracking	beat_track()	Beat positions in the audio signal
Onset Detection	onset_detect()	Transient event detection

Statistical Functionals

Apply to any time-series feature via compute_functionals(): mean, standard deviation, min, max, skewness, kurtosis.

Architecture

audiofeat
├── temporal/      # RMS, ZCR, energy, attack, loudness, rhythm, ...
├── spectral/      # Centroid, rolloff, flux, MFCCs, chroma, formants, ...
├── cepstral/      # LPCC, GTCC, ERB cepstral, deltas
├── pitch/         # Autocorrelation, YIN, pYIN, Praat backends
├── voice/         # Jitter, shimmer, CPP, harmonic ratios, glottal flow
├── rhythm/        # Beat detection
├── stats/         # Statistical functionals
├── io/            # Audio loading, single-file & batch extraction, CSV export
├── validation/    # Praat comparison, gold-standard scorecard
├── standards/     # openSMILE eGeMAPS/ComParE wrappers
└── catalog/       # Auto-discovered feature catalog

How it works: Each feature function checks if librosa is available. If so, it delegates to librosa's implementation for bit-exact parity with the research standard. If librosa is not installed, a pure-PyTorch fallback computes the same feature. Either way, you always get a torch.Tensor back.

Installation

Python >=3.8 is required. We recommend creating a virtual environment first.

pip (from PyPI)

pip install audiofeat

From source

git clone https://github.com/ankitshah009/audiofeat.git
cd audiofeat
pip install -e .

With a virtual environment

# Option A: venv
python -m venv .venv
source .venv/bin/activate
pip install audiofeat

# Option B: conda
conda create -n audiofeat python=3.11 -y
conda activate audiofeat
pip install audiofeat

# Option C: uv
uv venv && source .venv/bin/activate
uv pip install audiofeat

Optional extras

Extra	What it adds	Install command
dev	pytest, black, mypy, flake8	pip install "audiofeat[dev]"
examples	matplotlib, librosa, soundfile	pip install "audiofeat[examples]"
validation	Praat/parselmouth backend	pip install "audiofeat[validation]"
standards	openSMILE eGeMAPS/ComParE	pip install "audiofeat[standards]"
models	ASR, diarization, VAD, denoising	pip install "audiofeat[models]"
full	examples + validation + standards	pip install "audiofeat[full]"

Quick Start

Extract features from a file

The simplest way to get started. This extracts all core features and returns a flat dictionary of summary statistics:

from audiofeat.io.features import extract_features_from_file

features = extract_features_from_file("path/to/audio.wav")

# What you get back:
print(features["f0_mean_hz"])        # Mean fundamental frequency
print(features["rms_mean"])          # Mean RMS energy
print(features["spectral_centroid_mean"])  # Mean spectral centroid
print(features["mfcc_0_mean"])       # Mean of first MFCC coefficient

Compute individual features

For fine-grained control, call feature functions directly. Every function accepts a 1D torch.Tensor waveform:

import torch
import audiofeat

# Load your audio (or use a test signal)
sr = 22050
waveform = torch.randn(sr * 3)  # 3 seconds of noise

# Temporal features
rms = audiofeat.rms(waveform, frame_length=2048, hop_length=512)
zcr = audiofeat.zero_crossing_rate(waveform, frame_length=2048, hop_length=512)

# Spectral features
centroid = audiofeat.spectral_centroid(waveform, frame_length=2048, hop_length=512, sample_rate=sr)
rolloff = audiofeat.spectral_rolloff(waveform, frame_length=2048, hop_length=512, sample_rate=sr)
contrast = audiofeat.spectral_contrast(waveform, sample_rate=sr)

# Cepstral features
mfccs = audiofeat.mfcc(waveform, sr)
chroma = audiofeat.chroma(waveform, sr)

# Pitch
f0 = audiofeat.fundamental_frequency_yin(waveform, fs=sr, frame_length=2048, hop_length=512)

# Voice quality
jit = audiofeat.jitter(waveform, fs=sr)
shim = audiofeat.shimmer(waveform, fs=sr)

# Every result is a torch.Tensor
print(f"RMS shape: {rms.shape}")
print(f"MFCCs shape: {mfccs.shape}")
print(f"F0 shape: {f0.shape}")

Load a real audio file

from audiofeat.io import load_audio

waveform, sr = load_audio("path/to/audio.wav", target_sr=16000)
# waveform is a 1D torch.Tensor, sr is an int

Aggregate over time with statistical functionals

from audiofeat import compute_functionals

rms = audiofeat.rms(waveform, frame_length=2048, hop_length=512)
stats = compute_functionals(rms)
# {'mean': tensor(...), 'std': tensor(...), 'min': tensor(...),
#  'max': tensor(...), 'skewness': tensor(...), 'kurtosis': tensor(...)}

Batch extraction to CSV

from audiofeat.io.features import extract_features_for_directory

extract_features_for_directory("audio_folder/", "output.csv")

CLI

audiofeat includes a command-line interface for common workflows.

audiofeat --help

Extract features from a single file

audiofeat extract recording.wav --output features.json

Batch extract an entire directory

audiofeat batch-extract audio_folder/ output.csv

Diagnose your environment

audiofeat doctor --audio-dir examples

Checks installed dependencies, verifies audio files are valid, and reports any issues.

Browse available features

audiofeat list-features
audiofeat list-features --format markdown --output FEATURES.md

Advanced Topics

For Praat validation, openSMILE integration, the gold-standard scorecard, and troubleshooting, see docs/VALIDATION.md.

For the auto-generated feature catalog, see docs/FEATURE_CATALOG.md.

Testing

pytest -q

With coverage:

pytest --cov=audiofeat --cov-report=term-missing -q

Contributing

We welcome contributions! If you have new features, bug fixes, or improvements, please open a pull request on GitHub.

Citation

If you use audiofeat in your research, please cite:

@phdthesis{shah2024computational,
  title={Computational Audition with Imprecise Labels},
  author={Shah, Ankit Parag},
  year={2024},
  school={Carnegie Mellon University Pittsburgh, PA}
}