A recipe contains all the configurations for a certain model trained and tested on a certain database. Every experiment has several components each with their own configuration file. A recipe contains the folowing configuration files:
- database.cfg: This file contains all the paths where data should be read and written. This is the only configuration file that is not included in the repository because it depends on your setup. You can find more info about the database configuration in the next section.
- features.cfg: This file contains the configuration for the acoustic features (e.g. mfcc, fbank, ...). You can find more information about feature computers here.
- coder.cfg: this file contains the configuration about how the semantic representation is coded into a vector. You can find more information about coders here.
- acquisition.cfg: This file contains the configuration for the language acquisitionm model (e.g. NMF). You can find more information about language acquisition models here.
- cross_validation.cfg: This file contains the configuration for the cross_validation. For example how many blocks you want to split your data in.
- train.cgf: configuration for training, e.g. the sections in the database.cfg you want to use for training.
- test.cgf: configuration for testing, e.g. the sections in the database.cfg you want to use for testing.
- structure.xml: the semantic structure file. You can find more information about this file here.
Default configuration files can be created for the coder, the acquisition model and the feautures. You can find more information about deafault configuration files [here](## Default configuration)
The database.cfg file contains one section per speaker with the folowing options:
- audio: location of the audio file containing a list of all wav files. Every line in the file contains the name of the utterance and a pointer to the wav file.
- features (writable): a directory where the features should be stored
- tasks: location of the task file containing a list the tasks you can find more information on task representations here. Every line in the file contains the name of the utterance and the string representing the corresponding task.
your database.cfg should look something like this:
[speaker1]
audio = /path/to/speaker1/wav.scp
features = /path/to/speaker1/feature/directory
tasks = /path/to/speaker1/tasks/file
[speaker2]
audio = /path/to/speaker2/wav.scp
features = /path/to/speaker2/feature/directory
tasks = /path/to/speaker2/tasks/file
...
Default configuration files can be created for the coder, the acquisition model and the feautures. These default configuration files can be found in the defaults folder in the folder where theses components are defined.
A default configuration contains default values for the configuration files, if they are not defined in the configuration file, they will be taken from the default file. The name of the configuration file is the lower case class name of the class it is meant for.