FrameNorm

The frame normalization to use.

Set FrameNorm to one of the following options:

Note: The frame normalization used during the identification stage must match that used during development.

By default, Speech Server uses UBM if the templates are using a background model.

This means that when Speech Server generates scores for each audio frame, the server takes the background model score into account and subtracts the background model score from all the speaker scores. This process helps to reduce frame score variation (in cases where some frames might score higher on all models, including the background, leading to unstable results). HPE recommends that you use the UBM option for open-set speaker identification.

If you set FrameNorm to NONE, Speech Server does not take the background model score into account, but instead uses the scores as they are. HPE recommends that you use this option only if you trained your templates without using a Universal Background Model.

If you set FrameNorm to SET, Speech Server takes the scores for all templates, and normalizes them between 0.0 (the worst case) and 1.0 (the best case). This process can increase recall, but can also introduce more false hits. HPE recommends that you use the SET option for closed-set speaker identification.

If you set FrameNorm to AUTO in the audiotemplatescore module, Speech Server uses the frame normalization that was used during development when you produced the scores for threshold estimation. If development was not performed and there are no thresholds, Speech Server switches to UBM mode, unless you set ClosedSet to True, in which case Speech Server uses SET mode.

Action: AddTask
Task: SpkIdDevel , SpkIdDevelStream , SpkIdDevelWav
Type: String
Default:

UBM in audiotemplatedevel module

AUTO in audiotemplatescore module

Example: FrameNorm=SET
See Also: FrameNorm (configuration parameter)

_HP_HTML5_bannerTitle.htm