5
$\begingroup$

Related to High audio frequencies in television broadcast

My cell phone samples at 48000 samples/sec so the FFT app I installed can only detect frequencies up to 24 kHz.

If anyone has an audio frequency analyser that can go higher, would you please use it on a digital TV advertisement for Amazon Alexa and see if they are embedding any sort of ultrasonic signal to tell the device not to respond to the keyword?

When it first came out it responded to the word Alexa on TV all the time and now it almost never does. My guess is an ultrasonic signal is being used to suppress the keyword recognition. No-one has talked about this online or in the press that I can find, so I'd like some confirmation if this is true. (I can't see any other way they would be doing it that doesn't involve sending samples to Amazon which they should not be doing since the recognition light is not activating at all for these events)

So my question is whether an ultrasonic signal is present in Alexa TV ads and if so, what is it composed of?

$\endgroup$

2 Answers 2

1
$\begingroup$

Most speaker/amplifier systems will have analog filters that cut off far below 24 kHz.

In fact, with 44100 kS/s being a very common sampling rate, you can be certain that you can't transport any signal at frequencies above 22.05 kHz; Nyquist doesn't only apply to the digitizing process, but also throughout the rest of the signal processing chain. Since digital spectra are periodic with the sampling rate, any digital-to-analog converter must have so-called reconstruction filters that cut off at half the sampling rate on each channel.

In other case, what your phone can't hear is something that your TV cannot emit, either. (At least not intentionally – analog filters, especially affordable ones, are far from perfect)

My guess that the software simply got better at determining the "intonation" of commands. Luckily, for all Latin- and Germanic-based languages, the imperative is not that hard to discern acoustically – if I had a guess how most of an infant's and a dog's understanding of human language starts, it's that – and you can tell someone commanding "Alexa, do my dishes!" from someone noticing it in the center of a sentence as object like "I can't trust Amazon enough to let Alexa listen in on confidential conversations."

Another relatively common thing is Steganography – in the case of Audio that's usually called Watermarking – the art of hiding info in a signal in a way that is recoverable, but not noticeable to the observer. It's been used for easily one-and-a-half decades now[citation needed] to track bootlegged copies of movie recordings etc.

$\endgroup$
7
  • $\begingroup$ I agree: relying on an ultrasound signal to survive the audio stream compression + tv loudspeaker system is probably too fragile a solution. Watermarking is more likely; or maybe Alexa is programmed to detect other phrases that appear in the ad. $\endgroup$
    – MBaz
    Commented Feb 27, 2017 at 14:27
  • $\begingroup$ "My guess that the software simply got better at determining the "intonation" of commands." ... I'll trust your opinion that it's unlikely to be ultrasonic, but if you had an Alexa, you'ld never suggest intonation analysis! The recognition is extremely unreliable and there is no way it is being that clever. In any case, the ads for the Alexa itself stress the commands as a normal person would. Not to mention the rejection takes place before the end of the "Alexa" wakeword. A barely audible signature akin to a watermark is more plausible but it would be nice to detect it and duplicate it. $\endgroup$ Commented Feb 28, 2017 at 5:04
  • 1
    $\begingroup$ MBaz - no, it's not using any wider context - I've recorded and played back the small part of the ad that invokes the wake word. Whatever is encoded (if anything) is done within the space of that one word. $\endgroup$ Commented Feb 28, 2017 at 5:06
  • $\begingroup$ You think intonation of "Alexa" would be harder to recognize than correctness of recognizing commands. Now, "Alexa" is one word, and depending on how you recognize, the imperative might be easier to recognize than individual words. $\endgroup$ Commented Feb 28, 2017 at 10:01
  • $\begingroup$ No, I'm saying I've had one of these for over a year and I've woken it up using many many different ways of uttering the wakeword - querulous, supplicantive, exhausted, questioning, depressed, whispering, and yelling in exasperation. If it relied on saying it in a commanding tone of voice I would have noticed by now. It recognises the keyword regardless of intonation. $\endgroup$ Commented Feb 28, 2017 at 17:02
1
$\begingroup$

reddit user 'aspyhackr' may have found the solution. In his post ( https://www.reddit.com/r/amazonecho/comments/5oer2u/i_may_have_found_how_amazon_prevents_the_echo/ ) he says:

"I noticed that the Amazon commercials usually do not trigger the device, or if they do, she only momentarily wakes before ignoring what is said. I did a little research tonight and found that the Echo, while it’s processing the wake word, searches the Audio Spectrum and if is significantly quieter in the area of 4000hz to 5000hz, she will not wake for the word. I achieved this by going on YouTube, and playing with a voice recording of the name in Audacity.

I found that when I analyzed the spectrum of them saying her name, the spectrums were significantly quieter in the range of 3000hz to 6000hz. In some of those recordings, those frequencies appeared to be non-existent. In others it appeared like she boosted the surrounding frequencies to make the Echo see a gap in the spectrum.

I then got an Audacity plug-in to allow me to do a band-stop filter. I found that when I took a recording of someone saying the wake word, and ran the plug-in centered on 5200hz at “Half an Octave.” (Beware, Dual monitor picture) My echo would not wake, even sitting right next to the speakers!"

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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