In normal operation we encounter a number of false voice commands. To inhibit this false voice commands, I have used in the Cyberon DSMT program the global settings: energy threshold = 200 , confi.reward=-10 SG Diff.reword= -10 and ending Silence = 0,12. I have read the page "https://tool.cyberon.com.tw/DSMT_V2/index.php?lang=en"
For some commands, I have put the command confi.reward on -20. Does this value overwrites the global confi.reward or must we apply both values?
Are there other actions to do in order to inhibit the false voice commands?
已解决! 转到解答。
Hi @RitaD,
As mentioned on AN13212 Chapter "4.1 DSMT tuning parameters", both the Confidence Reward and the SG Difference Reward can be used to tune recognizability of the commands, so decreasing them would also decrease false alarms. Adding garbage words (as described on chapter 4.2) is also used to decrease detection of false alarms.
Finally, the command-specific rewards do not override the global values. This AppNote also mentions that "The global and command rewards work to offset each other. In other words, the total reward is the sum of the both." It is not an overwrite, but a combination of both.
BR,
Edwin.
Hi @RitaD,
As mentioned on AN13212 Chapter "4.1 DSMT tuning parameters", both the Confidence Reward and the SG Difference Reward can be used to tune recognizability of the commands, so decreasing them would also decrease false alarms. Adding garbage words (as described on chapter 4.2) is also used to decrease detection of false alarms.
Finally, the command-specific rewards do not override the global values. This AppNote also mentions that "The global and command rewards work to offset each other. In other words, the total reward is the sum of the both." It is not an overwrite, but a combination of both.
BR,
Edwin.