Annotation File Formats

Annotations are symbolic data associated with the recorded signals of the speech corpus. See chapter for a discussion of the different principles, label categories, methods and tools.

The format in which these data are stored has to be included into the corpus specification. Since no widely accepted standard exists and since very often speech corpora contain a new form of annotation that has never been applied before, many speech corpora contain proprietary formats that were defined only for that special occasion. Some of these formats have become commonly accepted and have been re-used in other collections.

In this section we list some more or less standardized file formats for annotation data and outline their respective properties. These properties can be described by the following criteria:

• platform independence
• self-describing
• plain text vs. markup text
• availability of tools

Where appropriate, the annotation formats will be described in these terms in the following sections.

• SAM Format^4.11
  The European SAM project defined a combined signal and labeling format for data collected in the SAM projects. The signal files contain only the raw digitized data while the description files contain the signal specification data, meta data and labeling data (annotations). SAM description files are widely used, e.g. in the SpeechDat telephone speech corpora.

  Pros:

  • Description files and signal files are separate and can be stored and distributed independently.
  • Easy to extend by new labels
  • Does not require any special software.
  • Works on all hardware platforms.
  Cons:
  • Basically a free-form text format - very weak syntactic constraints
  • Does not include flat or hierarchical linking of different annotation tiers.
  Recommendation: Use SAM for simple speech corpora, like read or prompted speech by a single speaker.
• EAF (Eudico Annotation Format)^4.12
  EAF (Eudico Annotation Format) is an XML and Unicode based annotation format. It supports both time-aligned and symbolic relations between annotation tiers. New tiers can be added easily.

  Pros:

  • Description files and signal files are separated and can be stored and distributed independently.
  • XML: Easy to parse and easy to manipulate.
  • Does not require any special software.
  • Works on all hardware platforms.
  • Open format: definition of new tier types very easy.
  Cons:
  • No label types for events between words (yet).
  • No tiers with both symbolic and time links.
  • EAF events are always time-consuming (no points in time).
  Recommendation: No experience at BAS with this format yet. Probably the most powerful format.
• BPF (BAS Partitur Format)^4.13
  The generic annotation format used at the Bavarian Archive for Speech Signals (BAS). BPF is quite similar to the old SAM format but has no fixed syntax and semantics for the annotation tiers. A BPF file consists of a SAM compatible header part to store signal specs and meta data and an unlimited number of tier blocks that contain the annotations to the signal. Reference between signal and annotation is done via the physical time scale (direct reference) or via so called symbolic links between annotation tiers (indirect reference). The format is open in that new tiers may be added to the file following some basic rules (5 basic annotation types). BPF uses the very basic UNIX concept of filters and lines of code which makes it very easy to create, manipulate and search BPF files on all platforms that provide some basic scripting language.

  Pros:

  • Description files and signal files are separated and can be stored and distributed independently.
  • UNIX filter concept: Easy to parse and easy to manipulate.
  • Does not require any special software.
  • Works on all hardware platforms.
  • Open format: definition of new tier types very easy.
  • Integration of other modalities than speech.
  • Overlapped speech, points in time and non-speech events.
  Cons:
  • No hierarchical structuring of tiers.
  • No intelligent search structure.
  • No specialized software to manipulate or visualize BPF files.
  Recommendation: Easy to use and easy to transform into other formats.^4.14
• ESPS
  A very simple and widely used label file format for simple event or segment labeling developed by ENTROPICS and used by their signal processing toolbox X-WAVES. Unfortunately ENTROPICS was bought by Microsoft and the support and distribution of X-WAVES stopped in 1999. Nevertheless, many labeling or database tools still use this format, for instance the EMU system (see [4]).

  Recommendation: Comparable to SAM but without the header part.

• AGS (Annotation Graphs)
  Annotation graphs are a general and very powerful concept for representing all kind of symbolic information related to a speech signal^4.15. Although the annotation graph per se is not a file format, Bird and his colleagues developed an XML based file format AGS that may be used to store all different kinds of annotation graphs.