14

The majority of smart assistant brands, such as the Amazon Echo and Google Home, offer very little in the way of customisation for the wake word (the phrase you use to wake up the device so it listens to you).

For example, Alexa only offers three choices and Google Home only supports 'OK Google'. A lot of users seem to be interested in the idea of custom wake words, yet none of the major brands have added support.

Is there any technical reason for restricting wake word customisation, or is it simply a branding choice?

I've read about Google's motivation for using 'OK Google', which suggests the branding idea might be true, but it also seems that wake word recognition isn't very accurate, perhaps indicating a technical reason. Would anyone be able to clarify which factor is the main reason?

  • 2
    One thing to remember is that the processing to wake like that must be done in a very low-power, always-on domain -- there's often specialized hardware available for this task in some of the larger SoCs (especially those used in phones), and I assume that the filter coefficients / other parameters are carefully calculated and loaded into that hardware. I forget which SoC Echo runs on, but I imagine there's a similar filter present. The name of the class of algorithms is escaping me at the moment – Krunal Desai Jan 11 '17 at 1:05
12

Yes there are several reasons.

This blog post explains it about the Echo and the wake word Alexa. I'll summarize a bit.

Wake word recognition is done locally and in real-time. That limits the lengths of the wake word due to obvious processing limitations. Furthermore users don't want to recite a poem to activate the smart assistant. Thus, it has to be short.

It has to work with an almost 100% accuracy when called and also not recognized with that certainty when not called for. That makes problems and also makes for a minimum length for wake words. Amazon's choice to allow Echo is quite surprising since it's just two syllables.

If we look at the usual suspects, we have Alexa (3 syllables), Amazon (3), Echo (2), Ok Google (4), Hey Cortana (4), and Hey Siri (3). All industry giants almost agree on three being a good choice number of syllables.

Strangely enough the most wanted wake word "computer" has also three syllables and would match that requirement easily. It's also not trademarked.

As the blog—and reason—tells us, we totally want to avoid false positives. Let's have a look about how established the words Computer, Siri, Cortana and Alexa are. This is the Google book corpus from 2008.

Ngram computer outshines alexa and siri

Exactly, Siri and Alexa are virtually flatlining against computer and Cortana gives an error. Not found. Makes sense since the corpus is from 2008. To give us some more perspective why computer is a terrible wake word another graph.

ngram of computer versus other words

This Ngram shows the two most popular US baby names of 2016 (for currency), as well as Tom and Dave also flatline against computer. Queen, basketball and police manage to register properly. Anyhow this gives us an idea why Computer, Earl Grey, Hot hasn't been allowed so far. People use the word computer way too often.

One more thing about false positives. Alexa rhymes with virtually nothing one would say.

19 things that rhyme with alexa

Computer rhymes with 74 things.

  • 1
    ... and this explains why Scotty did not get to wake the computer. – Ghanima Jan 10 '17 at 19:34
  • 2
    "I'm going to sex a bird...No not you Alexa!" – Reinstate Monica Jan 10 '17 at 20:56
  • 1
    "OK Google" is four syllables ("oh kay goo gull"), not three, and a lot more than that many phonemes. – Monty Harder Jan 10 '17 at 21:30
  • 1
    Alexa is a frigging first name... I personally know 2 people with that name, one of which is a cousin. Echo is used as is in my language, And I often say "there's an echo" when there's an echo on the phone or something. And amazon is a river / old tribe's name that a guy in roleplaying like me say relatively often. They really messed this up. – Olivier Grégoire Jan 10 '17 at 22:26
  • 2
    I'm not sure I buy that explanation of why the word "computer" wouldn't work. The word itself is not the only thing being recognized. Being the first word of an utterance and being followed by a pause are also significant pieces of information that identify the command word. – Kevin Krumwiede Jan 11 '17 at 4:06
2

Is there any technical reason for restricting wake word customization

When assistant device is not in use, the application processor (I think ARM in case of Alexa as well as Google Home) is suspended and taken to lowest possible power state. The wake word detection is left to very power efficient DSP which listens to the ambient noise/voices and runs a algorithm to decide if there is a match to the wake word. If it finds a match with good amount of confidence DSP wakes up the ARM core to get going with rest of the processing.

Now since the goal is to be power efficient the DSP in question runs the algorithm as well as stores the template pattern on the on-chip memory rather than the main on-board RAM. This allows system to even take the DDR RAM to lowest power state.

Since the DSP has a number key things to do and very little on-chip memory the Assistant wake words are limited to a few choicest ones that can be matched by the algorithm with high degree of confidence.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.