FAQ

FAQ 1

Question: For a Fixed Trigger, upon recognition, we get wordID and finalScore. How is finalScore used?

Answer: The user does not need to use finalScore. That value is used for checking with the 'reference' simulation. The reference results can be obtained using Sensory's snsr-eval™ tool. Compare the recognition results from the DSP to the reference results to ensure that the recognition from the DSP performs as expected. finalScore should not be used to check for the confidence of recognition. Recognition is binary (either yes or no); check the value of wordID to determine if recognition occured.

FAQ 2

Question: How is svScore used?

Answer: svScore is used for Enrolled Fixed Trigger (EFT) and User Defined Passphrase (UDP) to verify the enrolled user. This ensures that a non-enrolled user will not be able to wake up the speech recognition process. In this use case, svScore is compared with svThreshold in the THF-Micro™ application. If svScore is below svThreshold, it does not match the enrolled user and the speech recognition process should not start. The value to set the threshold will depend if it is EFT or UDP; Sensory's partners usually run their own tests to determine the value of the threshold they want to use. Sensory recommends starting with 3277 (max is 8192), which corresponds to 0.4 out of 1.

FAQ 3

Question: For a 'wake-to-voice-search' use case, how can the application developer pass the recognition endpoint on to the voice search process?

Answer: Starting from THF-Micro™ version 8.0.0+, the application code no longer needs to call SensoryFindEndpoint after wakeword recognition. Upon recognition, the endpoint should be stored in the returned RecoResult structure from SensoryProcessData. The endpoint should be used by the application layer as the starting point to stream the data to the voice search process.