 Why hello YouTube and the interweb It's nice to see you again. I'm Joshua Montgomery the CEO of Minecraft AI and Today I'm in our Lawrence production facility. I'm doing and asked me anything About our future roadmap for the micro project. So behind me you can see Fun 3d printer where we print prototypes for the mark one actually Stepping off camera for a second. You can see here some of the latest and greatest prototypes of the mark two Where you know this this piece came off of the the printer behind me You see the internal structure and whatnot. This is the one of the outer pieces for the new mark two and then this right here is a the first mock-up of the resonating chamber for The speakers and then of course the the faceplate for it Those are printed back there on the back behind me. You can see bd Who's responsible for doing assembly and customer support for all the the Minecraft units? I think he's got 16 mark ones to ship today. So That makes it a good day. And then of course We've got a full wood shop here on site 3x the CNC laser cutters Fuse deposition printer in addition to the SLA PCB fabrication facilities that we can reflow Circuit boards and do a bunch of other work like that. So This is our office in Lawrence, Kansas This is actually where things got started the whole purpose behind the microp project on day one was to Voice enable this space which we've got distracted and haven't yet done But the idea being to be able to turn on the lights Turn on and off the air compressor be able to operate all the equipment control the video screen in the lounge over there all via voice and our goal was to To just make something that was ubiquitous that anybody could speak to and when we went looking to see if anybody would let us If there were any tech out there, let us do it You know what we found was Siri which? Typical of Apple products was closed off proprietary and we couldn't access it Google Assistant was in its early days And then actually we never even heard of Amazon Echo before we got started In fact when I used to go and talk about Minecraft to other other people What I would start the conversation with you know, it's it's this little this little speaker in it that sits in the corner and it listens to you talk and you know, it does stuff on your behalf and They would look at me like I was insane. I actually saw it article recently that it talks a little bit about Amazon and Some of their future plans, you know and one of the things that this author was speculating about was that Amazon was going to build a home robot To voice enable that you could you can not only have the voice and voice technology but also have the ability to to manipulate the local environment and Simultaneously, I was reading about a Silicon Valley startup that had recently failed That had spent a year pursuing financing to build a home robot Right that was once again voice enabled and did things in the home and I couldn't find a single VC in the valley They talked to 200 VCs That was interested in funding them and I imagine that's a really similar experience to what my friends at Universal computing up in Toronto or my friend Jonathan Nostrom And Ivy went through When two or three years before the Amazon Echo came out, you know, those guys were out hustling money The guys at the guys built the Ubi up in Toronto, which is you know kind of an early smart speaker And in both cases, I think that the company has struggled to get investors to see their vision And my guess is knowing Jonathan and having spent some time with him It wasn't Jonathan's ability to communicate the message that was the problem He's a 500 graduate very well-spoken very intelligent man My guess is that the VC on the other side of the table Probably didn't have as many brain cells as they needed to see the future the way that Jonathan saw it And as a result, you know, there were some pretty significant challenges surrounding fundraising and putting together the type of Financial resources that are necessary to do a project of this the scope and so you know our friends at Amazon did a Superbowl commercial at the beginning of 2016 where they really unveiled the echo in a serious way with Alec Baldwin as a representative and All of a sudden voices like the hottest thing in Silicon Valley and VCs all have a voice thesis Where they're talking about Making investments of voice although they're making in many cases crummy investments or in a lot of cases You'll see a firm that has a strong voice thesis that has made no investments in a voice company And so I kind of wonder if Amazon is going to roll out a robot and just kill it like ship Umpteen million units globally and all of a sudden, you know, everybody in Silicon Valley is going to have a home robotics thesis and the guys At the start up the wired was writing about it be like we were doing that like two years before Amazon where it was our financing So anyway, I have digressed that was me talking about Why people didn't understand it when we got started the whole voice enabled thing clearly now everybody understands it Which is why, you know, we've been really luck You know, we were in the right place at the right time with the right piece of technology And we very much have the right team in the right community and it feels really good to be Sitting in this chair today, you know looking around me and seeing the work that's being done on my craft seeing Advances in the technology. I got the first samples back from the new speech synthesis engine mimic to Model today. Actually, that's not true. I got a set of samples a few days ago There were really really bad It sounded like some kind of nightmare that you would hear in the movie with like voices whispering in the background Like it was crazy bad. It was so bad that we didn't share it with the community. They were really bad and And just a few minutes ago. I got the the response from the next next generation of training We've since the first first voices we got we we've had a couple lambda quad machine learning rigs that we've now deployed for Keith and He's been able to run a lot more steps and the new voice I know we were supposed to have him deployed on April 30th, but We we're not quite there yet We've had some trouble getting as much data as we needed as fast as we needed it But now we've got the data We've got four hours of audio four hours of voice and it sounds awesome When we get to like six eight ten twenty hours of voice with the with the boy Danny the voice artist The new British voice is gonna sound fantastic. It was really gonna sound good and unlike our competitors who have Trained men everywhere to yell at their female sports speaker Our smart speaker is male No misogyny here my cropped. So anyway, let's move on and talk a little bit about the roadmap coming up I actually sent both some investors who are looking at writing us a check for our series a As well as my team an update yesterday on milestones and what we're looking for From both the team and the community between now and August. So so I can walk through that by August Our goal is to have my craft be feature complete and by feature complete I do not mean like every skill be perfect I do not mean the having full coverage of skills, you know the way that the the proprietary stacks do Feature complete as a very different meaning for me for me feature complete means all of the basic core technologies of the microstack are working and The data feedback loops that make the machine learning grow in intelligence over time are Closed and many of you in our community have already reached into our precise wakeward tagger and have tagged some of the precise queries that we That we our community has been donating for those of you that haven't please take a moment create an account at home.microft.ai when you log in up at the top it says tagging if you click that it'll load the precise tagger and all it is is effectively Hot dog not a hot dog Maybe it's hot. I mean that's really what it is. So it's either hey my craft not a my craft or My craft, you know as as the wake, you know the wakeward the person who just said my craft or they said Minecraft or they said Microsoft If they if it's a total miss, that's not my craft If it's right on so hey my crap, boom you hit the right on button, but if it's half the phrase So hey Mike would get almost or if they said twice a my craft. It didn't work a my craft a second time and The both samples are in there. You'll mark it as almost as a duplicate. So it's super easy Anybody can learn to do it in like three minutes like kids have been spending time working in that But that's an example of what we're trying to build a machine learning feedback loop And there's a blog post on our site that if you haven't read it I'd encourage you to read how to build a strong AI and AI that interacts just like a person where people can't tell What they're talking to? The answer is building these machine learning feedback groups so that people use the system By using it they generate the data The community tags the data categorizes it adds the intelligence to the data It gets fed back into the machine learning group at machine learning software and then fed back to the community for use And so you create this cycle where you're creating tagging categorizing training, you know Feeding the technology back creating the data tagging and over time it gets smarter and smarter and smarter as more data Pours into that loop and the community engages with it. So Precisely our wakeward tagger is our first instance of that because it is such a simple model It's either a microp or not a microp or you know some partial version of that So the community's done that this morning and we were at 49,000 tagged utterances That means that's 49 other 49,000 instances of a microp that people have confirmed not once but twice is Accurate that's been fed to our model and if you log into that you can see the tag at the top. It says Progress if you click the progress button, it'll actually show you a chart showing how accurate the machine the wakeward tagger has become So anyway, so we're doing that for precise already and that's deployed and we're really super happy with the scenario In fact, we're happy enough that we've now Started to release some money to our partner workaround For those of you who don't know workaround is a company that sends work to refugees From Syria or Atreia other places around the world that have strike Yemen They have strife and conflict where people are leaving their homes in many cases they're going to countries where they're not allowed to work and So what workaround does is it creates work for? Folks who are in that situation where if they have a smartphone or they have a tablet or they have a laptop that's connected to the Internet They can do this work and get paid a fair living wage for where they're staying So they they're not being taken advantage of they're not working at the low market rates They're able to actually feed their families at the rates that the workaround is paying and for a company like us That allows us to send simple work like wakeward tagging to somebody who is in dire need so that that person can can earn a living and feed their family and Do a good deed out there in the world. So So anyway, the precise stuff is working well enough that we're now we're now releasing some of the work to To pay trainers the idea there is eventually there'll be a balance between volunteer and paid We're thinking it's right around 20,000 volunteer utterances Unlock the money that we need to pay to have 5,000 paid utterances done And by doing that, you know, we we get the community continues to be involved But it gives the community a strong reason to come tag queries They're not taking work away from people in need. You know, we need to have the community involved in the process But when we reach a certain point it allows us to to release some finances so that we can pay You know folks in need to tag our answers and create work for people in the AI economy. So that's super exciting But anyway, precise is working well. Okay, but it's only one of several feedback loops One of the other feedback loops we're building is called persona where we take missed queries So if somebody says what's your favorite color and the AI can't answer that would get fed to the persona engine The community then decides on an answer Orange, I don't know if somebody in the community makes a suggestion somebody else in the community validates it in the future There may be some kind of voting process with that. We're not quite sure that That gets fed back into the machine learning engine and the next time somebody asks the same question The AI is able to answer Today we're gonna be we're doing that for subjective things like favorite color or favorite sports team or whatever Eventually that'll also include objective data. So they won't build tools that allow you to go out and grab a data source from Wiki data for example So a user might ask how tall was Abraham Lincoln, right and the AI might not be able to answer That gets fed into persona Somebody in the persona community says oh wiki data has got a table that has the height of all the US presidents in it They plug that in and in the future the AI can answer not just for Abraham Lincoln But also for all the presidents of the United States and because we are building a international community also an international President so if you want to know how tall Mr.. Trudeau or mr. McCrone are Chris you can you can ask the AI that question And so anyway, so for a for me Feature complete means all of those loops are up and operational the community is generating data The community is volunteering to share that data The community is tagging that data and that data is being used to improve the machine learning algorithms that answer questions from the AI At the same time we need to have a very high quality voice And I just mentioned that I just got the first samples of the new mimic 2 and they're awesome And we need to have the tools in place so that we can do skills arbitration or skills Deact that Still skill disambiguation arbitration Amazon's calling it skill arbitration They're bigger their vocabulary will probably stick. I would call it disambiguation. Although that's a hard word to say The idea there being if I say Play white noise, right? There are two potential answers to that question. They're actually multiple right it might launch the Pandora skill and play My white noise playlist on Pandora. It might launch Spotify and play The band white noise or it might launch a white noise skill, right? And actually play, you know, white noise static through the through the speakers so I can sort of waves or whatever through the speaker so I can sleep so You know one of the things that we're looking to build is an auto test function that allows us to identify At a minimum identify when a phrase might be ambiguous across multiple skills And then begin building tools that disambiguate there now My guess is and not my guess like the answer to how to do that is by using machine learning algorithms Which get sent back into our into our feedback loop? We have a plan to get that done. It's it's on the on the development Roadmap somewhere probably after the August that we're talking about so Between now and August we have a few things we want to get done precise wakeward tagger We want to have it finished we want to be able to collect data tag data feed it back to the engine feed the engine back into the Devices and to the instances of my crop down there. Good news. It's done. It's finished We're halfway there to getting the hey my cropped wake we're done When we get that done We will go ahead and And We will go ahead and begin adding additional wakewards And BD, can you go kill that please or connect it to the interweb one or the other? Sorry, I just kicked a device over there That's in testing So we'll have a mycroft done shortly after that we will begin working on a new wakeward I'm open to suggestions right now. I'm thinking it's Trinity Because we we do have any Shweppy reading a female American voice and it'd be great to have a female American voice for Mycrofts back in Trinity, but obviously there's there's other opportunities there so The next item after after the precise wakeward tagging that we want to have complete is to have the deep speech engine be the default STT engine for the speech to text engine for the platform and With the reason we want to do that is right now. We're taking all the queries from the users We bring them into our server and then we aggregate them we mix them together Right and we send them to a third party to be tagged in that third-party Software features back in answer now. There's no way Using the network paths and the authentication paths that we're using to figure out that you are the source of that data Whatever random and user you are But you know if you say your social security number into the device That's going to get sent to the STT engine and of course anybody who's reviewing those logs Would have access to your ST your personal information and that's something that we want to protect And so the best way to do that is for us to move the speech to text process on to Mycroft equipment and then nuke the logs for people who don't opt in to share data so that we don't keep the data at all We've been working very closely with Michael and Kelly and the Mozilla deep-speech team in Berlin for Six months or a year now those guys have made fantastic progress if you haven't already go and visit their project at voice Mozilla.org You can contribute your voice to that product that project in like 10 seconds You know just read three phrases it go and your voice will be included in the training set You can also review other people's contributions about about it that they're accurate So that team is has built a speech to text engine that is Equal to or better than humans with some corpuses of texting Let me let me highlight that last phrase some Corporances of text so if a user is Using the technology in the same way that they contributed the data Which means sitting in front of a computer screen reading the text from the screen It is as accurate or more accurate than a human at interpreting that speech which is fantastic That is the first step The issue is if the user is using a smart speaker in a kitchen with their kids yelling in the background the TV on and They have a thick foreign accent Deep speech today doesn't have any data Like that to train on so that we can improve the technology so deep speech really relies on very very clean data and As a result it's not as useful as we would have liked it to have been for a smart speaker Because of course people use the smart speakers or they use the the Raspberry Pi in environments that have noise They have accents, you know, there's a variety of different things that impact the quality of the speech and so You know, we made deep speech available to developers March 31st like we told folks we would However to enable you have to go through the CLI the command line interface We will be providing shortly. I think an option for that, but because of the lack of training data It's not as good as the third-party STT yet. Now. The awesome thing is we just finished negotiations with A Mozilla in fact BD just stand and sent the contract over to Mozilla over to our team to Plumb in my crops back end to the Mozilla deep speech data system So when you opt in when you if you don't opt into the micro stack We don't keep squat right and if we do keep squat it will last we will keep that if we do keep data on you We will keep the data just as long as I don't know that we're keeping the data Okay, as far as I know we are not I've told the guys not to I haven't personally inspected the servers But I'm confident that they've implemented our our goal our They've implemented their instructions if for whatever reason we find a log somewhere or we realize that There's a buffer somewhere that's keeping data that shouldn't be it'll last just as long as we know We don't know about it, right? So if you don't opt in we don't keep your data that our goal is ideally perfect privacy for end-users if people opt in So you opt in you go check the little checkbox on the thing and say hey I want to share my data with the community and you begin using the technology those utterances Will get sent to the Mozilla deep speech team where they will be transcribed Verified and fed into the deep speech engine so that Use it it improves over time with use in the real world and as far as I know We're the only deployed technology that's using deep speech in a serious way And I hope that they find other partners because more and merrier and more data we generate the better for everybody in terms of building accuracy but But anyway, that's something that needs to be built So deep speeches available But the plumbing between mycroft and Mozilla has yet to be built the good news is we've got the negotiation and the legal out of the way so we can actually do it and Eventually that will not only be the English utterances, but ours is in in other languages So the next big milestone is we want to make sure that the natural language understanding engine so that's the engine that takes a phrase of text, you know turn the lights on in the kitchen for example and Converts it into an intent so and so the turn the lights on the kitchen the intent would be activate IOT skill Object lights location kitchen toggle position on and then it executes a piece of code that runs up the IOT skill and goes out and turns on the IOT device So the NLU engine, you know, that's a very simple intent You know we and that's called a known entity rules based system So it knows that an object is lights so location is kitchen that toggle position is on that has a list of those things So our adapt engine, that's how it works. We also have a petitions engine which is a machine learning engine which takes example phrases and Trains a dialogue engine to understand what those phrases mean effectively And so each skill within our framework has a long long list of all the phrases that are supposed to activate that skill and data abstractions to abstract the data from that That gets trained in a machine learning engine and then the machine learning engine spits out JSON for the intent So both of those things are working our developers and many cases are using bodacious in some cases. They're using adapt in some case I think that makes them But the piece that we haven't built the piece that's on the roadmap today is the piece that takes queries and Allows us from the real world people who are using the technology and allows us as a community to categorize them and Put them with the appropriate skill. So if I say turn the lights on in the kitchen course great I say make it brighter where I cook Okay, it has no idea what to do with that brighter is not an object cook It may be an object, but it's a it's an activity, right? It doesn't communicate to an AI that uses known entities what the heck it's supposed to do But if we feed it into pedacious and we feed enough examples like that into pedacious So make it darker where I park make it bright where I mobile on whatever it is Over time with enough phrases the machine learning engine is able to suss out exactly what the what the intent was from the user The thing is as a developer who's developing a skill today has to imagine all of the potential phrases the user might use which of course Developers have fantastic imaginations. Don't get me wrong But it's very difficult to have a comprehensive imagination right unless you're maybe Jim Henson and so the The piece that we want to build between now and August is the piece that allows us to automatically take mist intent So make it brighter where I cook that gets fed into the engine we're calling persona, okay, and somebody in the persona community grabs that and says Oh, this is what the guy meant he meant make turn the lights on in the kitchen so I'm going to grab the IOT skill and I'm going to drag this phrase over there however we built the interface and Next time that pedacious intent gets trained, you know It'll know what to do with the with the user's utterance and be able to go out there and turn the lights on the way the user Users intended so that is is one of the pieces that we're looking to get done is to build the plumbing that allows us to capture data from people who opt-in only from people who opt-in like foot stop that a couple times and Feed it to the to the community and then on to the machine learning engine so that the technology can improve So the first part's done adaptive pedacious have been in production for a while, but the second piece the plumbing isn't quite done So next item is the company's mimic speed synthesis engine. That's our new Text-to-speech so it text comes in audio comes out. We want to be a fully deployed We're going to deploy that with one default voice and one hate voice, okay, so you know As much as I would like to Richard Stallman my craft right free to everybody Free software means no price free software also means freedom to use it You know nobody's ever charged for anything It all should be given away because there's zero incremental cost. It's all knowledge Ideologically on the line of that right, but ultimately Beatty's got to pay his mortgage right, you know, we have a great team of full-time people We have a bunch of server infrastructure that needs to live We have to operate in the real world where money is a thing You know, there are those that would argue that patronage would be a great Great way to fund the project and you know what if you have a patron out there who's got 10 or 15 million dollars Which is what it would take to really take this to the next level. Please connect me with them You know seven email founders of my craft that AI will find me And I'm happy to reach out to that patron, but since patronage is few and far between in a world where money's not free we have to develop mechanisms that allow us to both be open and Make our technology available, but also pay the bills And so for us one of the pieces that we think has value to our community is enhanced voices So voices that sound great voices that might be celebrity voices were given some serious thought to that You know voices that might have an accent that you like if you want to talk to somebody from Georgia We can add a Georgia accent that you know a female Georgia voice and we'll call her Samantha, right? So there's so much some options there So when we deploy the mimic 2 engine, we're going to continue with the British male voice Partially because it sounds cool, but but actually there's a lot of ideology there the concept that You know in this new me-too world like in a world where we're actually starting to pay attention to women and their role in the world and And how they're treated by especially here in the technology space It just seems to me to be a bad idea to deploy a voice assistant where you yell at the one You know and in many cases, you know, I'd be very curious as to what Amazon's data set looks like surrounding people screaming at Alexa I'm gonna go ahead and say that there's some misogynistic nasty stuff in there, right? And so the idea that we would use a male voice And a British male voice like a bond to come back out of the speaker and punch in the throat if you're mean to it Appeals to me. So the default voice will continue to be British male Default name will continue to be my craft show ox older brother and the smarter of the two brothers For the foreseeable future Enhanced voices though will be available and we're starting with the female voice that's being generated by our design lead Derek's wife Amy She's been recording now for a couple weeks We're getting really close to having a comprehensive data set and when that's done. We'll be deploying a High-quality American female voice for people who who choose to become patrons of the project So that is on the horizon if you had asked me two hours ago how I felt about reaching this milestone between now and August I had it told you I don't feel very good about it And then my CTO Steve sent me a sample of the the new mimic model, which I mentioned earlier that Really made me feel wonderful about where we are with the speech that this is sounds really good. It's awesome And then finally You know by 8 6 so August 6th I've asked the team to build internationalization tools One of the things that we're doing next year is taking my crop global So if you look around our little shop here in Lawrence, Kansas, I mean, it's a beautiful facility It's large. We've got a lot of really smart and capable people working both here in Kansas City We've got folks as far away as Stockholm and far away the other direction is Melbourne but It only speaks English right and it really doesn't speak English broadly it only speaks Really Midwestern American English the same type of English. I I speak which I'm told people from England say makes me sound dumb So there we go It doesn't really fully understand the thickly accented English. So English from maybe India I don't know if Kathy has trouble down down under in Australia But my guess is it doesn't work as well down there as it does For us here because of course, this is where the data set for a lot of the STT engine came from and so the goal for This year and for this sprint and for milestones between now and August is To develop the tools that allow us to take my craft into other languages and those tools Are primarily text-based although we do need folks to start yelling at my craft in other languages Mozilla is going to be helping with this on the 15th of May. I believe If it's not the 15th of May with somebody, please correct me in the comments Mozilla is doing a global sprint around adding new languages to the deep-speech project If you have a language that you want to make sure survives Okay, so it's pretty clear that English is going to survive Mandarin and Cantonese will survive French German What but there are a lot of languages that have small populations now? Icelandic Afrikaner and a host of other other languages out there and in some cases the populations that are speaking these language are beginning come become very very very small and and So if you speak a language where maybe the population that speaks your languages a couple thousand people Maybe it's one town one community or one village The tools that we are building should help to capture that language for prosperity both to Understand that language to speak that language and through our intent parsing engine and eventually the dialogue engine To be able to speak to an AI that speaks that language To be able to speak your native language, which maybe a thousand people speak but also have that simultaneously translated into Mandarin or English or For one of the more Broadway spoken languages is definitely something that's on the horizon using the data separate building so If you have an interest in supporting international languages, I strongly recommend that you participate in the global sprint the 15th and then certainly Look for me next year when I come through your country So we can talk about how you might contribute there the international language tools will be a series of online tools that allow you to Translate from any language my crop speaks to any language So if the two languages you speak are French and Swahili Once we get the French data set bill you should be able to go from French to Swahili and begin building prompts and dialogues Begin reading phrases for the speech understanding Tagging phrases for the the speech to text and then of course, you know if the technology is going to speak your language We're going to need you know the new voices. We're doing them a four hours We might get it down less than that I think we're going to take the current their first voice up to maybe 20 hours of audio But the eventual goal is a very small thing so two or three hours of somebody speak You know speaking using specific phrases into the computer creating the samples we feed on the machine learning engine and boom Magic machine learning The the technology is able to to go from text to whatever language we happen to be supporting You know one of the things about big tech is Big tech is so big Right that in order for them to grow their business. They have to get into big markets And so, you know the last I checked and don't quote me But I think the Google STT engine spoke about 80 languages, right? There are more than 7,000 languages spoken globally and Google is supporting 80 of them because of those are the 80 biggest markets and and with the size of their Their market their market cap with the amount of revenue that they have to generate to be successful with the amount of revenue They have to generate to grow Building technologies in these markets. Just it doesn't make sense for right and their business They're they're not there to not be evil. They're there to make money, right and You know unless they convert to a B Corp, they're legally required to only pay attention to making money Yeah In our case, we're a bit of a different animal, you know I think the final destination for my craft is is a B Corp. So it's a public benefit corporation That really is here to support the community not to make money. I mean making money is important because you have to pay people and what not but That's not doesn't have to be the only goal of a company and B Corp So the B Corp structure allows you to do that and So as a company like that, it's more important to me to be all-inclusive To support everyone to build as many languages into the technology as possible And so if we can get to seven thousand one hundred global languages, I don't see any reason. We can't Let's do it. Let's do it as a community So the tools that we're looking to build by August will allow anybody who speaks initially English and whatever the other languages To begin translating the prompts to begin contributing speech to begin contributing phrases And start that process We'll probably build a scoreboard of some kind that shows, you know, in a scale of zero to a hundred how far along we are in each language And then the question becomes Who in Iceland wants to make Icelandic available as a language for Automated speech recognition and speech synthesis and wants to build a an AI that it represents ice I did an AI that You know writes a book sometimes during its life an AI that Understands the darker winter an AI that You know has the ferocity of Of a people founded by of a nation founded by Vikings, right? Because the you know, it only takes one or two interested people in that country to translate it from English to Icelandic and then the language and the persona and the the ability to Synthesize now language all become captured in in perpetuity right they become available for everyone because of course our data sets are open And as the machine learning stuff and improves in quality Those data sets will become more and more important as people use this technology and new and novel ways So we're looking to have those tools built by August August 6th is when we leave and actually Iceland's the first destination And then after that we'll be coming to a lot of communities around the world in terms of broad Cities that we're going to be visiting. We're visiting our friends in Reykjavik. We'll be visiting Amsterdam After that will be in Berlin then Paris then Madrid then Lisbon and Barcelona Athens Rome Then we're going to head across the equator down to South Africa to visit Johansford We'll be up in Mumbai and then on to Singapore Then we go down under and we'll be in in Sydney and Auckland Followed by a trip to Fiji to talk about Polynesian languages After that we're headed up to Taipei Seoul in Tokyo to talk about tonal languages and how we can support those and then on to San Juan Buenos Aires and Santiago To look at a different flavor of Spanish from Castilian Spanish and so and in each one of those destinations we'll be branching out to speak to Communities in the surrounding area. So when we're in Greece You know, we'll be able to hop across the GNC and talk to folks who might speak Czech or might speak Slovak You know the same when we're in In Paris, maybe we get out to London and visit some folks who speak Welsh and maybe as far as As Ireland, you know to look at Gaelic and so on and so forth. So So we're super excited to be taking the trip and the goal of the trip is to Number one we want to make sure that we have an international Community that is supporting the Minecraft idea globally that you know somebody who wants to deploy a voice assistant and lives in Moscow can build a voice assistant that speaks Russian Using all using tools that have been community developed and work extremely on other languages So that they can apply that voice assistant for their special use case The same as somebody who speaks Swahili You know who wants to build a voice assistant, maybe for a project in the mobile market in Africa We'll have access to that technology and we want to have a global community that's supporting that and engaged and working together You know one of the things in this world I think that divides us as language that we don't speak the same language and in many cases We don't those languages confined or in form our thinking and so having voice technologies that allow us to do some taste translation that allow us to collaborate and interact Using the most natural form of human communication Makes a lot of sense and so we're we're looking forward to building that that global community over the course of the next year Looks like I have a few more minutes. I might cut this off early because I'm feeling repetitive today but anyway So some of the skills we talked last time was skills development review and deployment process We met a lot of progress there. I think we could involve the community a little bit more But it as of 1802 Those of you who didn't get the email you definitely look if you've already blow the skill You need to do a little bit of rework to pull it into 1802 We're we've implemented auto test procedures to help it help prevent skills from colliding on various different intents and We're requiring new skills coming in to have a certain structure and a certain layout So that they all played nicely together. We've had some instances where skills were stepping on each other and it creates a bad user experience so Laying that process out and then you know basically setting up a stone It's important Clearly there's a community role to be played there and we actually I think we already are engaging the community in a lot of ways But that's been a big rock for for 1808 The custom wake words is a big rock actually we have that you know Once we're done with the first hundred thousand with a microp we move on to the next one And once that one's done we move on to the next one So if you've got a wake word you're excited about whether it's Samantha or Trinity or Jarvis Get in there and help us finish the hey my craft wake word and then raise your hand and say we want to build this next So custom wake words is a big piece for us by the by the time by August 8th and or August of this year It looks like we're on track for that back-end account and metrics management We made a lot of progress with that. So Chris Bayer joining us has been a huge boost So Chris comes from the bats trading platform bats is a stock exchange located in Kansas City He's got mad expertise and building high availability back-end systems and has already Reduced the cost of running our back-end by a factor of five. So awesome work Chris Chris will be working with Mateus and we may be making a couple of additional hires to to really flesh out the back-end system and to better track metrics For for exactly how people are using the technology Obviously, we'll want to be tracking metrics for folks who opt in and continue to foot stomp that Deep speech integration. That's actually done So you can use deep speech as the default engine even though it's still not as high quality as some of the third-party engines If you're more interested in privacy than quality mimic or the deep speech engine is where to be Certainly the best way to improve that quality is to hit opt-in and donate your data the more data with your Personal voice that's in the deep speech Data set the easier it will be for the AI to understand it So there's a really strong incentive, especially if you have a thick accent if you come from a social demographic That's not very well represented Get in there and contribute that voice stop on zilla.org put in two or three hours of audio It doesn't take that long you can do it in an afternoon and that data will get pulled into the deep speech engine All of a sudden magically you'll find that it understands your voice really really well If you rehearse the phrases before reading them on the screen So instead of like reading it like rehearsed a couple times and speak it naturally that engine will pick your your language up better in the future So that's an important trick mimic to integration. So mimic to the new speech synthesis engine. We're now getting a Samples back from that. So the next step will be to do to fully integrate that into the stack It's early man, you know, it wouldn't surprise me if they had that done by the end of month I'm certainly by the end of June. So that's a an exciting milestone and then Auto testing is the is the other big rock and I talked about a little earlier, but building unit tests throughout the software stack That's a professional software development thing I Would argue there are many people who would argue about this. I Would argue fast and dirty up front get it working and start getting data and figuring out how users want it before you build a comprehensive auto Unit testing suite. I think that probably there are watchers out there who just had like a heart attack That's not you build software It is when you're in startup, but maybe it's not how you build at a university Maybe it's not how you build it a big court, but it's how you build it when you've got to build something fast And that needs to work before you run out of money Okay, I think we're past that point like we have out of a financing We've been blessed by the support of the community. We've been blessed by the support of some patient capable investors And so building unit testing into the software is an important milestone or important goal for us to reach between now and August That allows us to deploy high-quality software and to identify problems before they become problems Last but not least You know, I think one of the important milestones between here and the 1808 release Is there really a technical milestone? It's a community milestone today today. We cross 17,000 users a micro That's a pretty awesome number. I mean the fact that that many people Are interested in creating an account are interested in interacting with the technology is phenomenal Between now and August and in the next three to four months I'd love to see that number break 25 times And so, you know, if you have a friend who's interested in voice who's been talking about their echo or is maybe building skills in the Amazon community You have a friend who works at a company that might be looking at doing some kind of a voice integration for a product If you have a developer conference you're speaking at you're excited about what we're doing here at my crop share our story like let people know Why voice is important the technology we're building here at my crop not the stuff at Amazon not the stuff at Google Not the stuff at Microsoft that stuff's all vertical channels They're going to continue to own that stuff within their corporation, right? Like they're building what they're building and they're doing a good job, right? Their software is not going to get integrated into everything Because my crop is open Because companies and individuals can use it the way that they want to use it Because they can customize it both the wake word and the the skills and then the user experience and the personality My crops going to get built into pretty much everything if it runs a Linux kernel or it runs embedded There's a really good chance that a decade from now. It'll have some piece of the my crop stack in and that's super super cool and so You here today who are working on the software who are have set through a one hour Asked me anything To learn more about where we're headed as a community You the person who you know, even if you just interacted with the stack a couple times on a on an Ubuntu desktop today. That's cool You are contributing in a significant way to something that has a significant role to play in the future of technology Share that help us to grow the community to 25,000 to 50,000 to 75,000 It's very much a more the merrier situation, you know, how many languages can we support? How many users can we support? Broadly, can we make this technology? How many people who can't read today can a voice-enabled microstack teach to read tomorrow, right? How many people who can't see today will have global access to the internet and access to information in a completely transparent way because of the microstack How many people? You know, we'll get in a self-driving car and speak to a mycroft enabled voice assistant To tell him hey, I'm headed to the hospital where I'm gonna have my first kid take me to the hospital Right a lot and and that's super super exciting and we're looking forward to having you in our community If you've already made a contribution, thank you If you have yet to make a contribution reach in even if you just go into Homed out mycrafted AI and tag one precise query just one That's a contribution and we'd love to have have you as a contributor to the project I'm gonna step off There was in theory gonna be some kind of a feedback mechanism during this thing where I'd be able to like live interact with the community, but they kind of Stranded me on a desert island here in Lawrence Bay, and they're all in Kansas City building awesome software So I'm gonna punch out. I really do appreciate you taking the time to pay attention to to our project today I really appreciate your contribution. I really appreciate you as the user. Thank you very much for tuning in And I'm looking forward to the next time we connect Thank you