audformat is implemented in the audformat.Database. Written to hard disk it is converted to a YAML file (Header), which contains information about the raters, annotation schemes and meta information. Actual annotations are stored across (possibly) multiple CSV files (Tables). Each table column is linked to a scheme and/or to a rater. Each table row is linked to a media file, or a specific segment in a media file.

Parts of a database stored in the audformat on the hard disk.




Meta information, schemes, list of raters


Table with files or file segments as index and columns holding annotations, such as speaker id, emotion ratings, …


Audio/Video files referenced in the tables

The connection between the header and the tables is highlighted in the following sketch:


Connection between header definitions and table entries.

The annotations stored in the tables can be accessed as pandas.DataFrame. The following sketch shows an example instance of a database:

Header and Tables

Example content of tables and there connection to the header.