EnviSounds Dataset

Category selection process included the following 4 steps. (1) Identification included a thorough review of environmental sound taxonomies to identify a list of possible categories and their organisational criteria. (2) Synthsesis focused on extracting common structural characteristics to select categories that were most consistently indicated across various taxonomies. (3) Specialization ivolved exploring representative subcategories and consulting specialistic sound taxonomies and lists of classes from otehr databases to identify sounds produced by the same source. For example, although cats or dogs were included in most previous studies, both animals have more than one vocalization type in their repertoire. To correctly identify species- or source-specific sounds, appropriate literature was consulted (e.g., sounds produced by animals – cats, Pandeya & Lee, 2018; dogs, Molnar et al., 2008; birds, Briggs et al., 2012; sounds produced by cars, Morel et al., 2012; Park et al., 2019). (4) Filtering refers to the process of selecting the most frequently occuring and the most characteristic sounds for each class. First, lists of ecological frequency of everyday sounds (Ballas, 1993) and familiarity ratings (e.g., Marcell et al., 2000; Hocking et al., 2013; Burns & Rajan, 2019) were consulted. For example, familiarity ratings (with lower ranks representing more familiar sounds) indicated the sounds produced by a dog (M=1.3), cat (M=1.22), horse (M=1.27) and cow (M=1.54) are more familiar than other animals such as donkey (M=2.6), bear (M=3.31), goat (M=2.04) or lion (M=2.42; a full list of familiarity ratings can be found in Hocking et al., 2013). Then, for each of those selected objects or agents, we selected the most characteristic sounds. For example, sounds of dogs included sounds of barking or howling, but not sounds of snoring or walking, which are not distinctive for a dog (i.e., other animals might be snoring or walking).

The labels were constructed to be consistent grammatically and as generically descriptive as possible. They are all linguistic phrases derived from the most common labels occurring in the reviewed databases. Given that neither nouns nor verbs are sufficient to distinguish between certain sound classes (e.g., a 'dog' label does not differentiate between barking and growling; at the same time growling might not be indicative of a dog; Saygin, Dick, & Bates, 2005), we opted to include both, noun and verb (or its derivative) constructions in our labels. Their grammatical complexity was kept uniform by transforming all phrases into ‘noun + gerund’ forms. A similar approach was successfully exercised in previous research (e.g., Saygin, Dick, Wilson, Dronkers, & Bates, 2003). The same syntactic frame was used to construct the labels for all the sounds.

Each sound file in EnviSounds was analysed with librosa 0.11. Features can be grouped into four categories: time-domain, spectral, cepstral, and quality. The extracted features are stored as JSON files and a separate CSV summarises each feature with descriptive statistics per sound file.

Time-domain features

• RMS energy — root-mean-square energy per frame, computed with librosa.feature.rms. RMS energy is proportional to the perceived loudness of a frame and is a standard measure of signal power over time.


• Zero-crossing rate (ZCR) — a rate at which the signal changes from positve to negative is computed with librosa.feature.zero_crossing_rate. High ZCR indicates rapid sign changes typical of noise-like or percussive content, while low ZCR reflects tonal, slowly oscillating waveforms.


• Temporal centroid — indicates where in time most of the signal's energy occurs. A low value indicates an impulsive sound (most energy at the onset), while higher values indicate more sustained sounds. This feature can be useful for distinguishing a single clap from continuous applause, for instance.

Spectral features

• Spectral centroid — describes the centre of gravity of the spectral energy of sound (i.e., its dominant frequency), computed with librosa.feature.spectral_centroid. A high centroid reflects a bright, high-frequency dominated sound (e.g. birdsong, squealing tyres), while a low centroid reflects more dull or bass-heavy sounds (e.g. thunder, engine).


• Spectral rolloff — the frequency below which 85 % of the total spectral energy is contained, computed with librosa.feature.spectral_rolloff(roll_percent=0.85). Rolloff captures the upper frequency extent of a sound and helps separate broadband noise (high rolloff) from low-frequency sources such as distant thunder (low rolloff).


• Spectral bandwidth — the weighted standard deviation of frequencies around the spectral centroid, computed with librosa.feature.spectral_bandwidth. Broad bandwidth is characteristic of noise-like sounds (rain, wind) while narrow bandwidth indicates tonal content (whistling, bird calls, car horns).


• Spectral flatness — measures uniformity in the frequency distribution of the power spectrum and is calculated as the ratio of the geometric mean to the arithmetic mean of the power spectrum per frame, computed with librosa.feature.spectral_flatness. Values near 1 indicate a white-noise-like, uniformly distributed spectrum; values near 0 indicate a peaked, tonal spectrum. This is a key discriminator between texture-type sounds (rain, wind, running water) and tonal sources (birdsong, horns, bells).


• Spectral contrast — The difference in spectral energy between peaks and valleys in seven sub-bands, computed with librosa.feature.spectral_contrast(default parameters; 7 bands × frames). High contrast in the low bands indicates strong harmonic structure with clear valleys between harmonics; low contrast indicates a more noise-like spectrum. The feature captures aspects of timbre related to harmonic richness across the frequency range.


• Spectral flux — is defined as the sum of squared frame-to-frame differences in spectral magnitude and describes sudden changes in the frequency distributions of sounds (i.e. its dynamic variation). High spectral flux indicates rapidly changing spectral content, as found in percussive or transient sounds (footsteps, hammering, thunder). Low flux is characteristic of steady textures (engine drone, rain, running water).


• Spectral kurtosis — describes the flatness of the spectral distribution around its mean. High kurtosis indicates a strongly peaked spectrum (tonal or impulsive content), while low kurtosis indicates a flatter, more uniform spectral distribution. Unlike spectral flatness, which measures overall uniformity, spectral kurtosis is sensitive to the presence of sharp spectral peaks.


• Mel spectrogram — the STFT power spectrum mapped onto 128 mel-scale frequency bands using librosa.feature.melspectrogram, then converted to dB with librosa.power_to_db.


• Mean power spectrum — the dB-magnitude spectrogram averaged across all frames.


• Log-Mel band energies — a coarser version of the mel spectrogram using only 8 broad mel bands, computed with librosa.feature.melspectrogram (n_mels=8) followed by librosa.power_to_db. The 8 bands provide an interpretable, low-dimensional frequency-band summary suitable for direct comparison across sound categories, e.g. low-band energy is high for thunder and engine sounds, while high-band energy is high for birdsong and squealing tyres.

Cepstral features

• MFCCs (Mel-Frequency Cepstral Coefficients) — the first 13 cepstral coefficients derived from the log mel spectrogram, computed with librosa.feature.mfcc. MFCCs are the most widely used timbre descriptor in environmental sound recognition. The lower coefficients capture the broad spectral shape (related to vocal-tract or source colour), while higher coefficients capture finer spectral detail.


• MFCC Deltas — First-order temporal derivatives of the MFCC trajectories, computed with librosa.feature.delta using a 9-frame window. Delta coefficients encode how fast the spectral shape is changing from frame to frame, capturing the temporal dynamics of a sound's texture. They complement the static MFCCs and are especially informative for distinguishing periodic textures (sawing, galloping) from stochastic ones (rain, wind).

Quality features

• Harmonic-to-Noise Ratio (HNR) — per-frame ratio of harmonic energy to residual noise energy, estimated via harmonic/percussive source separation. First, we decomposed the signal with librosa.effects.hpss, and then computed RMS energy for both components. High positive HNR values indicate strongly tonal sources (birdsong, bells); values near 0 or negative indicate noise-dominated texture-like sounds (wind, rain, crowd noise).