Hacker News Reader

Comment by dperfect | Hacker News Reader

One of the more impressive things about the Echo is its far-field mic. I've used a lot of different voice recognition technologies, but Echo blows everything else away in terms of being able to speak in a normal voice from the other side of the room and have it still understand you. The actual intelligence behind the service may be lacking (or at least behind some of the competition), but for a general-purpose home application, the mic (in my opinion) is a significant part of Echo's reliability and success.

View on HN

Yeah, Echo is a good listener. In fact, when I ran the video on the OP's site and he said "Alexa, what time is it?" my Echo in the other room woke up and told me the time :)

That's actually one of my concerns about voice-driven interfaces, absent some sort of voice print authentication or tactile confirmation button. Imagine standing up in a crowded room and shouting "OK Google, delete everything!", or "OK Google, recite the first thouand digits of pi!"

We used to do this at work. "Hey Siri! Call [person's partner]". Everyone turned their voice activated stuff off quite quickly after that.

There was a demo a while back where some researches created radio interference in headset cables to trigger Siri remotely.

It would also be interesting if you could create sounds that sound like commands to the algorithms, but are not easily recognizable for the user.

Something similar to your idea was done for images recognized by neural nets and posted to HN recently. https://codewords.recurse.com/issues/five/why-do-neural-netw...

That could be a fun programming exercise for one of these devices you can hook into Alexa, see if you can catch unauthorised voices quickly and effectively without badly impacting response times.

e.g. you could do it by buffering the request, checking and submitting once you're confident about authenticity, but then you'd be delaying submission and response, depending on how much you'd need to buffer to gain sufficient confidence that the individual is authorised.

This was a common way to troll Xbox players: If they used speakers, shout "Xbox Off" loud enough via the voice chat (in a multiplayer game) and their devices turned off.

https://www.youtube.com/watch?v=mWZLa4AnN5k for a more humourous application.

Isn't that exactly why iPhone / iOS will ask you to say 'Hey Siri' several times when you activate the feature? I assumed it was tuning the recognition to key on just [approx] that voice.

It's not "secure" in any sense, but it cuts down on false positives and perhaps some very juvenile hacking attempts. Any professional hacker will generate a close-enough voiceprint that you can't use voice recognition inherently (purely passively) as authentication any time soon.

In my limited experience using Echo at my parents house over the holidays I was distinctly not impressed by the Echo's abilities to understand voice from across the room. We often had to repeat ourselves, only louder and slower to get the Echo to understand us. We do not have strong accents.

Exact opposite for me, does a great job at recognizing speech but a horrible job providing answers to even simple questions.

> a horrible job providing answers to even simple questions.

If you think of it as a device that performs specific actions (i.e., plays audio, gives a weather forecast, sets a timer, etc.) via an voice interface, rather than a general intelligence that can answer random questions, you're going to have a better time.

THIS. It baffles me how people expect devices such as the Echo to understand every nuance of human language, then get annoyed when it can't answer your overly-complex or vague question. Voice interfaces for me are always reduced to clearly enunciated and simple commands such as "set timer 10 minutes" or "play album Anjunadeep 4". You'll have a much better time than saying "alexaplaymesomealbumsbyaboveandbyond"

Well, in fairness, Google does a much better job than Alexa at 'doing what I want'

Echo has gotten much better at simple questions, and by prefacing "Wikipedia" prior to a query, I find the results satisfactory.

As anecdata goes, I have an Echo. It was an impulse buy when it came on pre-sale, then didn't know what to do when I actually received. I've been using it for quite a while now and the voice recognition is amazing, but the software is not really there yet. I can tell it to set up a timer from the kitchen (which is around a corner from the device) without much of an issue, even with my, somewhat strong, Argentine accent.

Many of the commands it "fails" at is when it tries to "guess" what I might have meant to say because it can't find something in its database. For example, if I give it a wrong name for a song (say "Alexa play She's Leaving by The Beatles" it might play some song called "She's Leaving" by some other band because it can't find that song under The Beatles.

Everyone has an accent. Do you mean that you speak General American (i.e. the North East non-coastal accent)?

Of course you're technically correct, but I find it quite common for Americans to understand that "no accent" refers to General American, even if they're not aware of the term "General American."

It's not just pedantics. I live in the UK, and some people who speak something between Received Pronunciation and the BBC accent also think they have no accent. People can't hear their own accents and presume they don't have any.

But this is more than just my bias for my own accent (which is very close to GA). When I meet other Americans with other regional accents, they will often describe me as having "no accent."

This is what is going to make Echo the next big computing platform. The voice recognition just works. It's like being on the Enterprise.

That's how I felt when using Siri and Cortana the first time. "Wow! Voice recognition actually works now??"

If you haven't tried the latest rev of Echo software, it blows Siri out of the water.

Very specifically, you can speak in a conversational tone with Echo, vs barking commands at Siri.

I don't have experience with Cortana, but the Voice Recognition capabilities in the XboxOne is garbage.

I've used Siri on the iphone and Cortana on a Nokia Lumia 939 and they seem about equal. I'm sure the echo is better due to the superior mic.

Agreed. It's an array with 7 MEMS mics, which allows the Echo to do beamforming so they can essentially focus on the speaker while rejecting most of the other sound.

There have been advances in deep learning based speech recognition to reject noise - http://usa.baidu.com/deep-speech-lessons-from-deep-learning/ for instance, but clean audio source helps massively, especially if you are trying to talk from any distance instead of being right next to the mic.

I haven't seen open source libraries for state of the art RNN/LSTM based speech recognition, anyone know if they exist?

MEMS mics have been taking over electret mics recently, they are way smaller and have better SNR (at equivalent size).

I'm building a 4 mic array at the moment, tetrahedral/ambisonics mic for VR use, the mics themselves are very cheap. Mouser and Digikey stock them - http://eu.mouser.com/Sensors/Audio-Sensors/MEMS-Microphones/...

Haven't looked into the software/beamforming part yet - does anyone know if there are open source libraries for that?

Some MEMS mics can also do ultrasound range, opening up pretty interesting possibilities for cheapo LIDAR type echolocation with a mic array - but that's for the next project :)

One of the simplest beamforming algorithms are delay-and-sum. This explanation has example code in C: http://www.labbookpages.co.uk/audio/beamforming/delaySum.htm...

PS: interested in your project, anywhere I can follow it?

> I've used a lot of different voice recognition technologies,

You've probably not used the Kinect controller? It has a very sophisticated distant speech system.

I believe the Kinect has a 4-element microphone array. Not quite as good as the Echo's 7-element array but it uses the same idea.

I don't know about the specifics of characterizing beam-pattern, etc. I am pretty sure, Kinect 1.0 can be more directional in spite of having fewer microphones because the microphones are cardioid, as opposed to the uni/omni-directional ones used in the Echo. Also, some parts of the design of the Kinect speech processing pipeline is public. You can read about in Ivan Tashev's (of MSR) book [1], IEEE Signal Processing Magazine article [2], and some presentation videos he has online [3]. Also, see his publication list, a lot of papers are relevant to the Kinect speech pipeline http://research.microsoft.com/en-us/people/ivantash/publicat...

Also obligatory, why the downvotes?

[1] http://research.microsoft.com/apps/pubs/default.aspx?id=1022... [2] http://research.microsoft.com/apps/pubs/default.aspx?id=2012... [3] https://www.google.com/search?q=tashev+video+lectures&oq=tas...

> Normal voice from the other side of the room.

High end smartphones have a similar high quality mic. It's no problem at all talking or using e.g. GoogleNow.

How often are you talking to your smartphone from the other side of the room? Usually it's a few inches from your face.

In my kitchen and in my office, it lies on the table. It very practical to just speak, I can walk around and do other things at the same time.

Not to mention smartphones seem to be explicitly designed to not provide this use case. Where is my always listening Google Now? If I have to take my phone out of my pocket and unlock it, I may as well type the question myself.

It's called loudspeaker or speakerphone mode. Of course it's not the default mode if you speak next to your ear.

The Moto X and its successors have had this feature for a while. Your phone can be sitting on the other side of the room, completely off, and you can say a certain catch phrase to activate Google Now.

I can usually activate Google Now from a few feet away on my Moto X (2013) but it can't understand what I'm asking after that. I have to get much closer for it to work.

Makes sense I guess. The "listen for the trigger" bit is done locally on those Motos (I've got the 2014 variant) but really all it needs to do is keep an "ear" open for that one phrase so it's probably got a better margin, only helped by local processing. The bit of audio that comprises the query has to get compressed and interpreted on the server side (not to mention the fact that it consists of a much wider variety of possible content to be parsed) so that can't help. I guess a dedicated chip that's just constantly asking "did anyone say 'OK Google?" has an easier job than a remote process asking "wtf does that fuzzy bit of compressed audio mean?"

(obviously being silly and oversimplifying but same idea)

All recent Nexus phones have always-on hotword support. I just tested it from 20 feet ("OK Google, what's the weather like tomorrow") and it worked fine.

Reply on news.ycombinator.com