Development of the Database for Environmental Sound Research and Application (DESRA): Design, Functionality, and Retrieval Considerations

Environmental sounds are gaining prominence in theoretical and applied research that crosses boundaries of different fields. As a class of sounds, environmental sounds are worth studying because of their ability to convey semantic information based on complex acoustic structure. Yet, unlike other large meaningful sound classes, that is, speech and music, the information conveyed by environmental sounds is not linguistic, as in speech, and typically is not designed for its aesthetic value alone, as in music (e.g., [1, 2]). There are numerous practical and specific applications for environmental sounds, which include auditory rehabilitation for hearing aid users and cochlear implants recipients; nonlinguistic diagnostic tools for assessing auditory and cognitive deficits in prelinguistic children; noise control and design of acoustic environments; auditory icons in computer displays; sonification, the process of representing information with nonspeech sounds (see [3] for a recent review). However, the knowledge base still lags far behind that of the other major classes of naturally occurring everyday sounds, speech, and music. One of the major hurdles is the lack of a standardized database of readily available, free, tested, high quality, identifiable environmental sounds for users to work with. There are various resources for accessing environmental sounds, some of which were detailed in [4]; however, that paper noted several caveats for users who are looking for sounds on their own. Among them were redundancy in time and effort needed to find the necessary information (which may have been found by others before), and the possibility of idiosyncrasies or occasional unwanted "surprises" (e.g., clipping or artifacts) in otherwise suitable stimuli. To correct those may require the user to have sufficient expertise in several technical areas such as digital signal processing and recording techniques, which may not, in and of themselves, have any relevance to the goals of the intended project.

To remedy this, this paper relates some findings of an ongoing project on the part of the authors assisted by James Beller, programmer, the Database for Environmental Sound Research and Application (DESRA—website: http://www.desra.org/) which aims to collect, label, edit when necessary, norm, evaluate, and make available a large collection of environmental sounds comprising multiple tokens (exemplars) of a wide variety of common sound sources. The collection of sounds and development of the Web front end are ongoing. This paper will describe the structure and function of the database, which reflects and responds to the complex ways in which we think about and use environmental sounds.

In order to optimally design a useful multipurpose database for environmental sounds, it is necessary to have a fuller understanding of the nature of environmental sounds, what they represent for humans, factors in environmental sound perception, and how their perception may be similar or different for different listeners. This information will guide sound selection by database users: researchers, designers, and musicians. The ability to use experimentally obtained perceptual criteria in sound selection, in addition to a thorough description of technical characteristics of the sounds, constitutes a unique feature of the present database. Although what Gaver termed "everyday listening" [5] is a frequent activity, the nature of the experience has been remarkably underscrutinized, both in common discourse and in the scientific literature, and alternative definitions exist [6, 7]. This is changing, as the numerous articles in this volume will attest, but still even our basic understanding of environmental sounds has large lacunae.

Thus, unlike speech and music, there is no generally agreed upon formal structure or taxonomy for environmental sounds. Instead, there are several prominent approaches to environmental sound classification that have been advanced over the last several decades [5–7]. A major initial contribution to environmental sound research is contained within the framework of Acoustic Ecology advanced by Schafer [6] who advanced the notion of the soundscape as the totality of all sounds in the listener's dynamic environment. Further extended by Truax [7] in his Acoustic Communication model, speech, music, and soundscape (that includes all other sounds in the environment) are treated as part of the same acoustic communication continuum wherein sounds' acoustic variety increases from speech to soundscape, while sounds' rule-governed perceptual structure, temporal density of information, and specificity of meaning all increase from soundscapes to speech. Importantly, the Acoustic Communication approach also treats listening as an active process of interacting with one's environment and distinguishes among several different levels of listening such as listening-in-search (when specific acoustic cues are being actively sought in the sensory input), listening-in-readiness (when the listener is ready to respond to specific acoustic cues if they appear but is not actively focusing his/her attention on finding them), and background listening (when listeners are not expecting significant information or otherwise actively processing background sounds). The theoretical constructs of the Acoustic Communication model are intuitive and appealing and have been practically useful in the design of functional and aesthetically stimulating acoustic environments [8]. However, directed mostly toward more general aspects of acoustic dynamics of listener/environment interactions, as regards cultural, historical, industrial, and political factors and changes at the societal level, it is still the case that more specific perceptual models are needed to investigate the perception of environmental sounds in one's environment.

In his seminal piece, What Do We Hear in the World [5], Gaver attempted to construct a descriptive framework based on what we listen for in everyday sounds. He examined previous efforts, such as libraries of sound effects on CD, which were largely grouped by the context in which the sound would appear, for example, "Household sounds" or "Industry sounds." While this would be useful for people who are making movies or other entertainment, he found it not very useful for a general framework. "For instance, the categories are not mutually exclusive; it is easy to imagine hearing the same event (e.g., a telephone ringing) in an office and a kitchen. Nor do the category names constrain the kinds of sounds very much."

Instead, he looked at experimental results by himself and others [9–12] which suggested in everyday listening that we tend to focus on the sources of sounds, rather than acoustic properties or context. He reasoned that in a hierarchical framework, "Superordinate categories based on types of events(as opposed to contexts) provide useful clues about the sorts of sounds that might be subordinate, while features and dimensions are a useful way of describing the differences among members of a particular category." Inspired by the ecological approach of Gibson [13], he drew a sharp distinction between "musical listening", which is focusing on the attributes of the "sound itself", and "everyday listening" in which " the perceptual dimensions and attributes of concern correspond to those of the sound-producing event and its environment, not to those of the sound itself."

Based on the physics of sound-producing events, and listeners' description of sounds, Gaver proposed a hierarchical description of basic "sonic events," such as impacts, aerodynamic events and liquid sounds, which is partially diagrammed in Figure 1. From these basic level events, more complex sound sources are formed, such as patterned sources (repetition of a basic event), complex sources (more than one sort of basic level event), and hybrid sources (involving more than one basic sort of material).

Gaver's taxonomy is well thought out, plausible, and fairly comprehensive, in that it includes a wide range of naturally occurring sounds. Naturally there are some that are excluded—the author himself mentions electrical sounds, fire and speech. In addition, since the verbal descriptions were culled from a limited sample of listener responses, one must be tentative in generalizing them to a wider range of sounds. Nevertheless, as a first pass it is a notable effort at providing an overall structure to the myriad of different environmental sounds.

Gaver provided very limited experimental evidence for this hierarchy. However, a number of experiments both previous and subsequent have supported or been consistent with this structuring [12, 14–18] although some modifications have been proposed, such as including vocalizations as a basic category (which Gaver himself considered). It was suggested in [16] that although determining the source of a sound is important, the goal of the auditory system is to enable an appropriate response to the source, which would also necessarily include extracting details of the source such as the size and proximity and contextual factors that would mitigate such a response. Free categorization results of environmental sounds from [16] showed that the most common basis for grouping sounds was on source properties, followed by common context, followed by simple acoustic features, such as Pitched or Rumbling and emotional responses (e.g., Startling/Annoying and Alerting). Evidence was provided in [19] that auditory cognition is better described by the actions involved from a sound emitting source, such as "dripping" or "bouncing", than by properties of their causal objects, such as "wood" or "hollow". A large, freely accessible database of newly recorded environmental sounds has been designed around these principles, containing numerous variations on basic auditory events (such as impacts or rolling), which is available at http://www.auditorylab.org/.

As a result, the atomic, basic level entry for the present database will be the source of the sound. In keeping with the definition provided earlier, the source will be considered to be the objects involved in a sound-producing event with enough description of the event to disambiguate the sound. For instance, if the source is described as a cat, it is necessary to include "mewing", "purring", or "hissing" to provide a more exact description. There are several potential ways to describe the source, from physical objects to perceptual and semantic categories. Although the present definition does not allow for complete specificity, it does strike a balance between that and brevity and allows for sufficient generalization that imprecise searches can still recover the essential entries.

Of course sounds are almost never presented in isolation but in an auditory scene in which temporally linear mixtures of sounds enter the ear canal and are parsed by the listener. Many researchers have studied the regularities of sound sources that can be exploited by listeners to separate out sounds, such as common onsets, coherent frequency transitions, and several other aspects (see, e.g., [20]). The inverse process, integration of several disparate sources into a coherent "scene", has been much less studied, as has the effect of auditory scenes on perception of individual sounds [21–23]. As a result, the database will also contain auditory scenes, which consist of numerous sound sources bound together by a common temporal and spatial context (i.e., recorded simultaneously). Some examples are a street scene in a large city, a market, a restaurant or a forest. For scenes, the context is the atomic unit for the description.

Above these basic levels, multiple hierarchies can be constructed, based on the needs and desires of the users, which are detailed in the next section.

The structure and functionality of the database are driven by the users and their needs. The expected users of this database are described below.

These values can be automatically calculated for a sound upon entry [2].

Sounds will be uploaded using the screen shown in Figure 4. The sounds accepted for uploading will all be high quality—at least 16-bit 22 kHz sampling rate for wav files, at least 196 kbps per channel bit rate for mp3s, with little or no clipped samples. The sounds must be recordings of physical sources—no synthesized sounds will be accepted. The sounds can either represent single sources or scenes. This will be designated by the originator upon uploading and verified by the admin. If the sounds represent single sources, a determination will be made as to the isolation of the source, that is, whether only the source is present or whether there are background sounds present.

There are four main front end functions that are essential to the functioning of the database. They are user enrollment and access; uploading sounds to the database; the search engine; and, perceptual testing.

8.2. The Search Engine

The portal to accessing sounds is the Sound Omnigrid shown in Figure 2. When users first enter the database, they are presented with the Omnigrid plus options for searching on and managing the sounds.

If the user selects "Search" a search screen will come up. Users will be able to search upon any of the sound data and/or keywords. For example, a search on the keyword "rooster" returned the screen shown in Figure 3.

Furthermore users can specify multiple criteria on any of the fields (e.g., search for sound types "horse" and "train" in either mp3 or wav format), and the number of tokens returned for each criterion, allowing users to easily create sound lists for further use. Where multiple sound tokens fit a particular criterion, a random one will be returned (or several random ones, if multiple tokens are requested) and this will be noted in the sound's search history so that usage statistics can be calculated and to prevent the same sound tokens from always being used. The users will be able to save the search criteria and the sound lists returned as part of their user profile. The users will also be able to organize sounds into selection sets to use in the experiment module of the database program (not discussed here) or for downloading and managing the sounds.

A current feature of many music search engines is "Query by Example" (QBE) in which a user can search for a song that "sounds like" something, either a selection from another song or by some descriptive terms, such as "light and romantic." One example is the Shazam application for the iPhone which can recognize a song based upon a sample submitted to it [41]. It would be useful to apply this paradigm to environmental sounds, so that users could search for a sound that "sounds like" a sample submitted to it (e.g., if a user had a car crash sound they liked and wanted more that sound like it, they could retrieve those) or to identify a hard to identify sound sample based upon returned matches.

However, extending the technology that is currently used in most Music QBE searches is problematic for environmental sounds. Most musical QBE searches do an encoding of the song signal using some compression algorithm, such as Mel Frequency Cepstral Coefficients (MFCCs) or projections onto basis functions, which is the MPEG-7 standard. The compressed version is compared to stored examples and the closest match returned via a distance metric, a common one being a Gaussian Mixture Model [42, 43], which is one of the options in the MPEG-7 standard [44, 45]. These programs are greatly aided by the fact that nearly all musical examples have a similar structure. They are harmonic, which makes encoding by MFCCs particularly effective, extended in time, continuous (not many silent intervals), and nearly all have a few basic source types (strings, percussion, wood winds, and brass).

Environmental sounds, on the other hand, since they are produced by a much wider variety of sources, basically encompassing every sound-producing object in the environment, are much more varied in terms of their spectral-temporal structure, some being continuous and harmonic (a cow mooing), some continuous and inharmonic (wind blowing), some impulsive and inharmonic (basketball bouncing), and some impulsive and harmonic (ice dropping in glass). Finding a common coding scheme that can encompass all of these has proven quite difficult, and most systems that classify and recognize music well do quite poorly with a wide range of environmental sounds [46, 47]. It should be noted that this refers to individual environmental sounds in isolation. When multiple sound sources are combined in a soundscape the envelope tends to be smoother, and the long term spectrum approaches a pink noise [48, 49]. In this case, algorithms used for musical classification also perform quite well [50].

In musical QBE systems it is often the case that a musical sample is first associated with a certain genre, such as "rock" or "classical" due to gross acoustic characteristics common to members of that genre. Some algorithms for environmental sounds will similarly initially classify a sound clip based on a semantic taxonomy and then use signal features to narrow the search. An example of this is Audioclas [40, 51, 52] which uses Wordnet semantic classifiers [53] and was incorporated into Freesound.org's similarity search procedure. However, in [40] it was reported that the probability of retrieving conceptually similar sounds using this method was only 30%. An alternate taxonomy put forth in this issue by [54] is based on the one formulated by Gaver described earlier. A comparison of the two can be found in the Roma et al. piece in this issue [54].

However, there is another way to restrict the search space and thus enable better automatic recognition of environmental sounds which uses only signal properties. In [16] strong correlations were found between the ranking of sounds in a multidimensional similarity space and acoustic features of these sounds. For example, sounds that were grouped together on one dimension tended to be either strongly harmonic or inharmonic. A second dimension reflected the continuity or discontinuity of the sound. Based on this finding, a decision tree can be proposed for automatic classification of sounds, as shown in Figure 5. While this does not cover all the sounds, it is a fairly simple structuring that does account for a large percentage of the sounds necessary for an effective classification system and would greatly enable the development of a true automatic classification scheme for environmental sounds.

The structure of a database of environmental sounds has been outlined, which will relate to the way people listen to sounds in the world. The database will be organized around the sources of sounds in the world and will include a wide variety of acoustic, contextual, semantic, and behavioral data about the sounds, such as identifiability, familiarity, and typicality as well as acoustic attributes such as the spectral centroid, the duration, the harmonicity, semantic attributes of sounds, such as "harshness" or "complexity", and details of the recording, for example the location, the time, the distance of the microphone from the source, and the recording level, along with a citation history. A flexible search engine will enable a wide variety of searches on all aspects of the database and allow users to select sounds to fit their needs as closely as possible. This database will be an important research tool and resource for sound workers in various fields.