Example

As an example appendix C contains the specification for the WebCommand speech corpus, while in appendix D you will find the original main corpus documentation file. Based on these data a validator and a client might come up with the following `validation contract' specifying the reference and the validation check list:

Validation Contract

between .... (validator)

and .... (client)

1. The client and the validator agree that the validator will perform a validation of the speech corpus `WebCommand' (corpus) with regard to common rules of best practice in the field of SLR production. The corpus and the software tools that were used for the annotation will be provided by the client (or a third party such as the producer) including all specifications and documentations as being originally delivered to the client.

2. The validation process is based on the specification of the corpus (technical annex ...). The validator will deliver his results in a confidential report to the client. The report should address errors / deviations from the specification in such detail that the client is able to correct these errors (if possible).

3. The validation will cover the following:
   • formal checks for completeness, terminology, readability and parsability of signal files, meta data and annotation files.
   • check for superfluous files in all locations of the SLR.
   • check of the technical specifications of signals files; empty signals; clipped signals; corrupt signal files
   • speaker distribution as stated in the specification. Documented sex checked in 50% randomly selected speakers.
   • completeness of documentation
   • consistency of speech corpus with documentation
   • readability (on Windows and Macintosh) of documentation files
   • manual validation of a minimum of 10% randomly selected transcription files including the adherence to the specified prompt texts. Errors to be reported are: typos (the `Duden' being the reference), mismatch between transcript and the spoken utterance in the recording, wrong noise marker.
   • formal check and completeness of the lexicon (coverage), check for mismatch to spelling in the transcripts.
   • manual validation of a minimum of 15% randomly selected lexical entries. Errors to be reported are: typos, mismatch to spelling in the transcripts, inconsistent canonical pronunciation.
   • readability of distribution media on Windows, Macintosh and Linux
4. Time plan of validation:
   Begin: ....
   Intermediate report: ....
   Final report: ....
5. Compensation
   ....
6. Confidentiality and legal stuff
   ....

Signature Validator Signature Client