store_slice: Provides the ability to store a multidimensional feature set...

In humlab-speech/superassp: A speech signal processing using various framworks using a wrassp-like interface

Provides the ability to store a multidimensional feature set related to a part of a signal.

Description

This function allow you associate a feature vector of high dimensionality to the part of the signal from which is computed. The part can be a time slice (with defined start and end times), or the entire signal. The collection of time windows from which data has been obtained is viewed as sparse collection of slices. A slice defined by its start and end times. The key features of the idea of sparse collection of slices is that even though the set of features extracted for the slice collection is consistent, a consistent spacing of slices against the time line of the signal file is not required. This design choice helps with the storage of output of computationally intensive procedures that summarize a whole recording or just the part of it where it makes sense to compute it.

Usage

store_slice(
  mediaFileName,
  values,
  measureNames,
  start_sample = NULL,
  end_sample = NULL,
  fileExtention = "sli",
  outputDirectory = NULL
)

Arguments

mediaFileName	The signal file from which the set of values has been obtained. Primarily, this will be the name of the speech signal file, but all signal files handled well by the read.AsspDataObj and can return a sample rate and number of samples will likely work.
values	A list of values to be stored. The identity of each measurement value should be indicates as informative list names, as it will help retrieval of values. If the list values are not named, they will be identified by an integer value instead. The first time the user stores values from an analyse in a particular sparse slice collection file will determine what features may be stored and what names may be used for them.
start_sample	The first sample of the signal file that was submitted for analysis, and for which the results should now be stored. If NULL, all samples from the first sample to the end_sample will be included.
end_sample	The last sample of the signal file that was submitted for analysis, and for which the results should now be stored. If NULL, all samples until the end of the media file will be used.
fileExtention	This file extension will be used when making the sparse slice name from the mediaFileName. Defaults to ".sli".
outputDirectory	The directory where the slice file should be stored. If not defiled (NULL), the sparse slice file will placed in the same folder as the media file.

Details

The primary example of a feature set belonging to a sparse collection of slices is a large collection of acoustic measurements of voice obtained from just the portion of the recording where the participant produces a maximally prolonged vowel. Computing the output may be computationaly intensive, and applying it across the entire signal just makes no sense for the application. As an example of this situation, please refer to the vat function which takes approximately 60x the duration of the sample to complete, and is only valid for vowel productions. The participant may also produce more than one prolonged vowel in a recording, and there is therefore no unified analysis window to apply to the entire speech file (such as what is the case when constructing a spectrogram from a set of equally spaced spectral slices). The specification of start_sample and start_sample for each slice additionaly allows for partially overlapping slices in flexible way, and can be added to the sparse collection iteratively. The user may, for instance, apply the analysis to the entire vowel, and then just to a 2s portion starting 1s into the vowel, and store the results in the same datafile for later use.

An example of a whole file feature set would be a high-dimensional acoustic description of an entire speaker or recording. This set is not well suited to be aligned to a particularportion of the speech signal, and in this case start_time and start_time is not set (NULL). Even though the the feature set is alined to a maximally large portion of the speech signal, it is still considered a slice in the sparse collection of slices.

Sparse collection of slices are flexible, but the implementation of them assures that two important features are upheld at all times:

1. The results stored in a sparse collection of slices must be identially shaped. That is, they may only hold the features defined when you first insert data into them. It is of course possible to have multiple sparse collections (with unique file extensions) where the user may store outputs of different analyses associated with a single speech signal.

2. Two slices may not be associated with exacly the same portion of the speech signal. If the user supplies a new result to the same time slice (including a slice including the entier file), the old values will be overwritten. This is to ensure that the user will always be able to deternine which set of output values is obtained from the just the start and end times supplied to the get_slicedata, and also assures the most efficient retrieval of the data. If the user needs to, for instance, apply the same analysis multiple times to the same portion of the signal file, then the user should instead use multiple sparse slice collection files.

humlab-speech/superassp documentation built on May 8, 2024, 2:27 p.m.