PipeSound

The PipeSound audio repository is a collection of audio files related to domestic in-pipe acoustic events.

The archive includes:

A weakly-labelled Soundbank containing the real source recordings,

A strongly-labelled Synthetic Dataset composed of 2k synthetic audio observations,

A strongly-labelled Synthetic Soundbank containing the atomic audio blocks of the Synthetic Dataset.

The Soundbank comprises the following classes of low-reverberated signals:

Backgrounds*,

Disturbances*,

Dish Washers,

Showers,

Sinks,

Taps,

Toilets,

Washing Machines.

The special classes Backgrounds* and Disturbances* are obtained from the remaining classes but linked separately(1). The background recordings are also used as noise references to post-process the Soundbank and improve the signal-to-noise ratio of the source recordings. The count of the Soundbank class instances is reported in the related Summary.

The Synthetic Observations composing the Synthetic Dataset are obtained as a combination of the augmented artificial audio blocks, which are stored for reference in the Synthetic Soundbank. The augmentation is obtained by recombination of the source recordings and through additional transformations, such as time-invariant pitch shift and pitch-invariant time stretch. A bespoke acoustic model simulates the in-pipe reverberations accounting for straight elastic pipes filled with inviscid water.

The audio files are equipped with a set of descriptive time/frequency domain figures and strong annotation metadata. Annotation metadata are stored using the .jams format extended with a bespoke .json schema named pipeSound. A comprehensive description of the audio items is stored in the related annotations, where the link between corresponding audio blocks is also fully maintained.

All the class instances of the Soundbank are roughly evenly represented in the Synthetic Soundbank. The Summary of the dataset reports the related count and illustrates this property. 

The unprocessed source audio recordings are stored in .dat format and were acquired using a processing chain composed of:

Brüel&Kjær type 8103 hydrophone,

Brüel&Kjær Nexus 2692 charge amplifier/ gain amplifier/ anti-aliasing filter,

ADLINK USB-1210 16bit ADC.

Copies of the unprocessed recordings and the related filtered audio samples are stored in the Soundbank in the uncompressed .wav format. The Synthetic Dataset and Synthetic Soundbank signals are stored in the compressed .mp4 format to minimise the space required(2). The scale factor to convert the dimensionless signal into the equivalent pressure signal in Pascal is stored in the comment of the audio file metadata (multiply the normalised amplitude by the scale factor). The associated Matlab software library can be downloaded from Github(3).

(1) The .lnk files are generated for Windows machines and might need conversion in the case of different operative systems.

(2) The full uncompressed version of the dataset can be requested at [e-mail].

(3) https://github.com/pipesoundlibrary