Improve Performance

Some common situations can reduce the performance of speaker identification.

Training and test data differ in audio quality Try to use training data for each speaker that closely resembles the recording conditions and quality that you expect in the files that you want to process.
Unbalanced speaker thresholds

If HPE IDOL Speech Server has too many false positives, use a higher Bias value when you finalize the speaker score thresholds (in the SpkIdDevelFinal task).

If the false negative rate is too high, use a lower Bias value.

Insufficient training or optimization data HPE recommends that you use at least five minutes of data for each speaker for training, and more if possible. A similar amount of data is required for each speaker for the development tasks if you are training score thresholds for open-set identification. If you use less data, this can compromise the general performance of the speaker templates.
Insufficient ‘unknown’ data used during optimization This situation can lead to poor speaker thresholds, which in turn lead to excessive false positives.
Very short audio segments are used in identification Short audio segments can reduce the accuracy of speaker identification. Use the MinSpeech and MinNonSpeech parameters to specify the minimum size of audio segments, and the DiscardShort parameter to set the minimum segment size before discarding the result.

_HP_HTML5_bannerTitle.htm