audiofeatures

A lightweight Python library for audio preprocessing and handcrafted feature extraction for Machine Learning, Deep Learning, Speech Processing, and Music Information Retrieval (MIR).

---

Overview

audiofeatures provides an end-to-end pipeline for extracting handcrafted audio descriptors from music, speech, and environmental sounds. It performs preprocessing, window-based analysis, and returns the extracted features as a Pandas DataFrame ready for machine learning workflows.

---

Why audiofeatures?

• ✅ 185 handcrafted audio features
• ✅ Built-in preprocessing pipeline
• ✅ Window-based feature extraction
• ✅ Dataset-level feature extraction
• ✅ Pandas DataFrame output
• ✅ Machine Learning ready
• ✅ Supports WAV, MP3, FLAC, OGG and M4A
• ✅ Simple, clean API

---

Installation

pip install audiofeatures

---

Quick Start

Single Audio File

from audiofeatures import extract_features

df = extract_features(
    audio_path="song.wav",
    window_size=0.5
)

print(df.head())

Entire Dataset

from audiofeatures import extract_dataset

df = extract_dataset(
    audio_folder="dataset/audio",
    window_size=0.5,
    output_csv="features.csv"
)

print(df.head())

---

Feature Categories

Category	Description
Preprocessing	Cleans and standardizes audio before feature extraction
Temporal	Describes signal characteristics in the time domain
Spectral	Describes frequency-domain properties
Cepstral	MFCC-based spectral envelope descriptors
Tonal	Harmonic and musical pitch descriptors
Pitch	Fundamental frequency statistics
Rhythm	Onset and rhythmic activity descriptors

---

Feature Descriptions

Audio Preprocessing

Mono Conversion

Converts multi-channel audio into a single channel for consistent processing.

Resampling

Converts audio to a common sampling rate (default: 22050 Hz).

Normalization

Scales audio amplitudes to a consistent range without altering relative dynamics.

Silence Removal

Removes silent regions to focus analysis on meaningful audio content.

Audio Cropping

Allows extraction from a selected region of an audio file.

---

Temporal Features

RMS Energy

Measures the average signal energy within a window. Higher values generally indicate louder audio.

Zero Crossing Rate (ZCR)

Counts how often the waveform crosses zero. Useful for distinguishing voiced, unvoiced, and noisy sounds.

Energy

Represents the total signal energy in the analysis window.

Peak Amplitude

Maximum absolute amplitude observed in the window.

Crest Factor

Ratio between peak amplitude and RMS energy. Indicates impulsive or transient signals.

Dynamic Range

Measures the difference between the quietest and loudest portions of the signal.

Energy Variability

Measures how energy changes across short frames inside the window.

Amplitude Variability

Measures variation in waveform amplitude over time.

---

Spectral Features

Spectral Centroid

Represents the "center of mass" of the spectrum. Higher values correspond to brighter sounds.

Spectral Bandwidth

Measures the spread of frequencies around the spectral centroid.

Spectral Rolloff

Frequency below which most (typically 85%) of the spectral energy is contained.

Spectral Contrast

Measures differences between spectral peaks and valleys across frequency bands.

---

Cepstral Features

MFCC

Captures the spectral envelope using the Mel scale. Widely used in speech and music analysis.

Delta MFCC

First-order derivative of MFCCs describing short-term spectral changes.

Delta-Delta MFCC

Second-order derivative of MFCCs describing spectral acceleration.

---

Tonal Features

Chroma

Represents the intensity of the 12 musical pitch classes independent of octave.

Tonnetz

Captures harmonic relationships between pitches for tonal analysis.

---

Pitch Features

Pitch Mean

Average estimated fundamental frequency.

Pitch Standard Deviation

Variation of pitch over the analysis window.

Pitch Minimum

Lowest detected pitch.

Pitch Maximum

Highest detected pitch.

---

Rhythm Features

Onset Strength Mean

Average strength of detected note or beat onsets.

Onset Strength Standard Deviation

Variation in onset strength.

Onset Strength Maximum

Strongest detected onset within the window.

---

Supported Audio Formats

• WAV
• MP3
• FLAC
• OGG
• M4A

---

Returned Data

The library returns a Pandas DataFrame.

For extract_features():

• One row per analysis window.

For extract_dataset():

• One row per analysis window.
• Includes song_id and filename columns.

---

Applications

• Music Emotion Recognition
• Speech Emotion Recognition
• Music Information Retrieval (MIR)
• Environmental Sound Classification
• Audio Event Detection
• Bird Sound Recognition
• Feature Engineering for Machine Learning

---

Roadmap

• Parallel feature extraction
• Additional audio descriptors
• Feature visualization utilities
• Real-time streaming support
• GPU acceleration (optional)

---

License

MIT License

---

Author

Dhyan Sudheer

Integrated MSc Data Science
Amrita Vishwa Vidyapeetham