Snowboy can not make a distinction between "turn on" and "turn off"

[78226415]

Hi, There is a hotword:turn on, you would see it by click the link https://snowboy.kitt.ai/hotword/899. When building your personal model to the hotword, the first step is to record your voice, then test the model, if you say:turn on, there should be a Ding voice and you can save the model by click the "save" button. But, if you say: turn off, there also is a Ding voice and you can save the model by click the "save" button. Why? Snowboy can not make a distinction between "turn on" and "turn off"? Go on next test. My test is based on the snowboy-android-64bit examples, you could get it from the GitHub: https://github.com/kitt-ai/snowboy, select the "android-64bit" branch, and build it. When you run the app, there should be a Ding voice after you say: alexa. After download the model: turn_on.pmdl, copy it to the directory: assets/snowboy/, then replace the code "alexa.umdl" by "turn_on.pmdl" in the file:Constants.java. Built the app and run it, click the "Start" button, there should be a Ding voice after you say: turn on, but, there should be a Ding voice after you say: turn off. I did another test, chinese hotword, built the hotword:"声音大点" model, Snowboy can not make a distinction between "声音大点" and "声音小点", you will find the same result as the test above. Is there some body has the same or similar issue? Is someone could help me? Thanks

[chenguoguo]

Those are usually called false triggers or false alarms. The personal models are not very robust, and it's likely that those models will trigger similar voices/sounds.

[keshavramaswamy]

Sorry, if I am hijacking this thread (happy to open a new issue if required) - I face a similar issue with my personal model as well. I use RaspberryPi as well - the model does not detect anyword with sensitivity < 0.7 and has a high false alarm rate otherwise. Is this a known issue with all personal models as they are only trained on a small amount of voice samples ( 3 in this case) ? So the best way to improve the model's detection is to train it on a larger pool of voice samples then? Also I use this usb microphone (have set audio_gain = 10): https://www.amazon.com/Kinobo-Microphone-Desktop-Recognition-Software/dp/B00IR8R7WQ do you reckon a better one is recommended to improve the detection?

[chenguoguo]

@keshavramaswamy did you use the same microphone for recording and testing?

[usarawgi911]

@keshavramaswamy I know this is quite old a thread, sorry for the same. But you can possibly help me with the following: I am using the same microphone as you with raspi zero for hotword detection with snowboy. And i randomly get one of the two errors :

Any help will be really appreciated. Thank You