Voice search - impact on recognition rate

In my experience, when the new feature “voice search” is activated, voice recognition rate (in challenge/response mode) goes down the drain, to the point of being unusable.

Thanks for your feedback.

The first comment is if you are not in the US or Canada, for now I would disable the Voice Search from the Talk to Mira menu. It adds a lot of time indexing words you will never use. When we roll out the full international support there will be options to only generate airports in your region.

The second comment is currently a few small percentage of users use the advanced Challenge/Response mode. Most use the default behavior of saying CHECK or CHECK ITEM and SKIP or SKIP ITEM where Mira first reads both the challenge and response. See below for my explanation of single syllable commands and you will always get better recognition saying CHECK ITEM and SKIP ITEM, although for some such as myself CHECK and SKIP never give me issues. Depending on your accent that may vary. We have tested extensively with the standard commands with full Voice Search and have not found significant issues.

With your case of using Challenge/Response mode and you defining all of the commands there are multiple factors that may affect it. When Voice Search is activated, the vocabulary of words the voice recognition engine has to deal with goes up a considerable amount. It is indexing all of the names of airports currently in the US and Canada as well as spelling them saying letters and with a phonetic dictionary (alpha, bravo, charlie, …). If some of those names happen to be similar to words you may be setting as a response then there can be less accuracy. The quality of the microphone also can effect it. The outside mic on an iPhone or iPad will not perform as well as testing with a very good noise cancelling mic like a Bose A20. When creating responses it will always help if the response have two or more syllables. Apple, Amazon and Google figured this out and it is why their wake words are Alexa, Siri and Ok Google. Single syllable words will be more of a challenge and have more error rate than two syllable. Three syllable will be even better. Your accent when you pronounce words can also affect it. All this said, I ran through a good portion of your preflight checklist with Voice Search enabled and this was the result. This was just using the iPhone mic which again would be improved better with a good mic. There are no smoke-and-mirrors, I just launched your checklist and started recording the screen.

I will say before recording the video I did run through and there were a few commands that didn’t work the first time which always ended up being the very short single syllable ones. When I said the command a second time it did work.

Voice recognition has improved considerably over the years but it is still not perfect in all scenarios to all audiences especially trying to be used in a harsh environment like a cockpit. These type of workarounds can help mitigate issues.

Long term we are also going to look if we can put the airport info dictionary in a separate layer from the normal commands where you first would need to say an initial command to enable that dictionary. That way the voice recognition engine would have a more focused dictionary to deal with when navigating your checklist compared to asking for airport info.