Comments to WebCommand

Two major errors occurred during the production of the WebCommand speech corpus that caused the production costs to be about 25% higher than estimated. Both errors might have been avoided if the production had adhered strictly to the standards recommended in this cookbook.

The first error was a very reduced in-house pre-validation after the setup of the recording hardware and procedures. Because of this sloppy procedure it went unnoticed that the same prompts were displayed in both recording rooms. According to the specification however, the prompts had to differ. This error was not even detected in the final validation but was discovered by the client after the first delivery. It is very likely that a proper pre-validation by an external partner (or the client) might have detected this logistical error much earlier thus saving a considerable amount of manpower and money (see section for details about pre-validation).

The second error was of a technical nature. Although the pre-validation was performed on both recording setups in the two recording rooms and showed valid recordings, it turned out in the final validation that in one of the recording rooms the input selection was later unintentionally set to another channel and all recordings were in fact empty. A number of speakers had to be contacted again and asked for an additional recording session to fill up the missing data. This error could easily have been detected very early if quality control in form of random tests on the recorded data had been performed (for details about quality control refer to section ).