Yandex holding opened for users mobile devices access to your voice assistant Alice. As you know, there are already such assistants in Russian - Siri, for example, but the scope of its use is limited to iOS devices. Or "Ok Google."
We decided to test these applications and asked all three the same questions.
Alice jokes and cheeks
To begin with, it turned out that you can talk to Alice only through a smartphone, the tablet has not yet been connected. Simulates live dialogue, recognizing natural speech (spoken and written), and gives answers to questions (saying them aloud and displaying them on the screen). According to the creators, Alice is the world's first virtual assistant, which is not limited to a set of predefined responses. Works in the Yandex application for iOS and Android, as well as in the beta version of the voice assistant for Windows. Almost artificial intelligence.
Alice speaks in a Russian voice Scarlett Johansson - speech synthesizer uses specially prepared recordings of the actress Tatiana Shitova, dubbed in the Russian box office most of the roles of the actress, including the role of virtual assistant Samantha from the movie "She". Shitova's voice is familiar to the Russian audience also from the dubbing of other foreign actresses.
Many of Alice's personality traits are defined by a set of phrases composed by Yandex editors. One of the authors of Alice's character was a journalist and writer Vladimir Guriev (which she does not admit). However, the developers emphasize that Alice is not limited to a set of predefined editorial responses: the assistant's neural network is trained on a large array of Russian-language texts, including network dialogues. This affected the nature of the program: some users are faced with the fact that it refuses to answer questions or is insolent. Alice's creators constantly monitor her behavior and correct it. She knows how to respond emotionally: for example, depending on the context, she can show cheerfulness or sadness. We launch the application.
- Alice, you read " Regional newspaper", Which is published in Yekaterinburg?
- Of course.
- What's the last thing you read?
— « Great expectations» ( There was no such title, no material - ed . ).
- I'll look now. ( Submits to several articles on this topic - ed. ).
- When do you need to change the shoes of the car?
- I'll look now. ( Again refers to articles - ed. ).
- It depends on what kind of life.
- Alisa, what will this winter be like in the Urals?
- I'll look now. ( A reference to forecasters - ed .).
- Alice, say hello to Siri.
- It's a shame to hear!
Useless interlocutor
"Ok Google" is a long-time assistant in finding information on the Internet, an already familiar technology voice search... Can you work with him in a dialogue mode? Let's try:
- Google, do you read the "Regional newspaper", which is published in Yekaterinburg?
- So, here's what I managed to find on the Internet on this topic ( referring to our site and, nicely, to my last material - ed. ) .
- Why should vodka have a strength of 40 degrees, and not otherwise?
- So, here's what I managed ... ( repeat about what we managed to find - standard start each answer and referring to the corresponding article - ed. ).
In questions about a car, winter and life, Google, like a mantra, repeats the first phrase (boring, you can die) and sends you to the appropriate sites. And when asked to say hello to Alice from Yandex, he sends us to a page with a recipe ... how to make a chocolate candy for a hamster.
The interlocutor from Google, obviously, is none, but the reference book is not bad.
Siri on your mind
Siri - cloudy personal assistant and a question-and-answer system, the software client of which is included in iOS, watchOS, macOS, and tvOS Apple... The app uses natural speech processing to answer questions and make recommendations. Siri adapts to each user individually, learning his preferences over time. Positioned as artificial intelligence.
The voice for the first version of Siri for American users belongs to the voice actress Susan Bennett. In an interview with CNN, Bennett said that in 2005 she signed a contract with ScanSoft and spelled incoherent phrases on a voice recorder for four hours a day for a whole month. In 2014, dubbing appeared in Russian.
By the way, communication with Siri is a humorist writer Mikhail Zadornov dedicated one of the monologues. When he suggested to Siri for directions famous expression"And you went to ...", she sent him to Ukraine.
- Siri, do you read the "Regional newspaper", which is published in Yekaterinburg?
“I'm afraid I don’t know ...
- Why should vodka have a strength of 40 degrees, and not otherwise?
- Wow. We have a problem. Try again. ( So she didn't say anything - ed .).
- When do you need to change the shoes of the car?
- V currently we are not going anywhere. If you want to get a route - tell me where.
- How to live your life interesting?
- Everything is fine!
- Siri, can you say hello to Alice from Yandex?
Conclusion: if this is artificial intelligence, then I am the Pope.
So what do we have in the bottom line? More or less Alice is drawn to the interlocutor, everything else is ordinary reference books, and do not expect anything more from them.
- Published in # 190 on 10/12/2017
- On October 10, 2017 Yandex is officially its voice assistant Alice.
- At the heart of "Alice" - neural networks that recognize speech, interpret it, compose responses and synthesize the assistant's voice. "Alice" learns on huge amounts of text to communicate with the user on free topics.
- There are also "editorial" lines written by the team to answer popular questions.
- One of the main difficulties in developing such a system is to understand what users will be interested in and how they will ask about it. Someone asks the search engine "where am I", someone - "what is my geolocation."
- The main differences between "Alice" and competitors (for example, Siri) - integration with its own search engine company and the presence of a neural network, with the help of which the assistant independently generates new replicas.
Why Yandex took up the development of a voice assistant
By 2017, the world's largest IT companies presented their voice assistants: Siri from Apple, Assistant from Google, Alexa from Amazon, M from Facebook, Cortana from Microsoft work in the English-speaking market. Asian - Duer from Baidu and Bixby from Samsung.
The company started to create "Alice" at the end of 2016. The service began to be developed because the entire industry is moving in this direction: the user wants and is ready to solve his problems in the form of a dialogue with virtual assistant, the algorithms are ready for this - and you can go from voice input to meaningful dialogue.
When the user has the opportunity to ask a question by voice, he begins to speak in a human way. Instead of “Lada Kalina 2007, a cheap cardan shaft to buy” asks: “Where can I buy a cardan shaft for a 2007 Lada Kalina”? Does not make up search query, does not formulate it, as he used to do it, but simply asks.
In this direction, move in parallel and search algorithms, and dialog interfaces. But in the voice assistant there is a binding to the context - that is, the ability to retain the meaning of the remarks and understand whether subsequent statements are connected with them.
Let's say you asked "Alice": "What is the weather tomorrow?" - she answered you: "+7 and it's clear." The next question may be: "What about the weekend?" It is easy for a person to understand it, but it is difficult for a machine - after all, it does not even have keyword"weather". The Dialogue Assistant knows how to match the lines and understand that you are still talking about the weather.
Ilya Subbotin
Product Manager "Alisa"
According to the developers, Yandex users already know how to access the company's services by voice - the Listen, Yandex command works in Search, Navigator, and other applications. So they won't have to get used to the new interface.
Ilya Subbotin, head of product "Alisa"
From a user perspective, the project team says, with the advent of Alice, a few things are changing. In addition to the dialogue interface, the assistant has acquired a personality - "Alice" shows emotions, knows how to joke and "does not tolerate humiliating treatment." You can now refer to her by name: "Listen, Alice."
From the technological point of view, a multilayer neural network is now responsible for the work of the assistant, which learns from arrays of texts, understands the meaning of the request and synthesizes speech on its own.
How Alice works
"Alice" is built into search application Yandex. You can communicate with the assistant both by voice and by entering requests from the keyboard. If possible, "Alice" answers the question asked directly in the dialog interface, in other cases it shows search results upon request or the desired application.
Under the hood
At the heart of "Alice" are neural networks: they recognize speech, analyze it, determine the topic of the dialogue, highlight useful information for searching, and synthesize the assistant's responses.
Almost all calculations, besides voice activation (recognition of the phrase "Hey, Alice"), are done in the cloud. “Of course, it would be interesting to transfer part of the calculations to a smartphone, so that Alice can sometimes work without the Internet,” says Ilya Subbotin. "But the computing power of smartphones is too low for such tasks."
Technologies machine learning have reached the level where the user can simply speak with the application and see that it understands it - and the synthesized voice acquires a sound close to natural.
Alice's team
At the same time, the world scientific and technological community is only at the beginning of the path - the algorithms will be improved further. Speech recognition, synthesis, definition of the subject of the conversation - all this should work even better.
There are general product metrics that can be used to evaluate the assistant's work - user satisfaction, frequency of application use. But there are also more complex parameters, explains Subbotin. For example, the ability to be in the context of the conversation: the assistant must remember what the user asked him about, be able to correctly interpret additional questions and give relevant answers.
“From the relevance of one specific dialogue in the future, it will be possible to move to personalization and multitasking: ideally, when I go to work, I can ask what to wear,“ Alice ”will look at the weather, the route to the office, decide what is the best way to get there, and based on this, he will advise me what to wear. "
It is not yet clear how to objectively assess the system's ability to conduct a dialogue. The length of the dialogue says little about its meaningfulness: the conversation can drag on if the user is interested in discussing something with the assistant or if the assistant cannot solve his problem for a long time. But even a short dialogue does not always indicate a successful solution to the problem.
Search focus
The developers of Apple's Siri voice assistant are deliberately increasing the number of conversational scripts - so as not to show the user search results. Alice's team is not faced with such a task, because Alice is a different product.
Apple doesn't do this out of choice. It's just not a search company. We are completely different. For us, search in the first place is best product which we have.
It's too early to guess how an assistant integrated with search will be adapted to work on other devices, says the head of the Alice product: the industry has not yet figured out which devices will be in demand among the audience, and in addition, such devices will imply different interaction scenarios. for which the system can be adapted.
“In any case, the user has a task that he wants to solve - by interacting with a voice speaker or, say, a car control system. In the case of the speaker, the emphasis will probably be on music or news - and we have music and news services. In the driver's car, he is interested in maps. It is unlikely that he will ask the car: " plastic windows buy low"".
Large tech companies are betting on smart device ecosystems to enhance the interaction with the assistant. If a device without a screen needs to show search results or open an application to answer a question, the assistant can launch required service on another - for example, on a smartphone.
“It's just a matter of adaptation to the device,” continues Subbotin. But, to be honest, we don't know anything yet. One of the industry leaders - Amazon - is still experimenting with formats: the company has an Echo speaker, a mini version of Echo Dot, options with a camera, with a screen. "
Personality
The personality of "Alice", according to the developers, is one of her main competitive advantages... Projects aimed at an English-speaking audience do not try to create a strong emotional connection with the user, but, on the contrary, are rather removed from him.
To give an assistant a personality, the team needed to determine the gender, age, name and character traits, describe the behavior and make up the lines that the assistant would use to dialogue with the user - “editorial” responses.
Together with a journalist, writer and former head by the Yandex search marketing group, Vladimir Guriev, the developers described the character's character and selected suitable cues for him.
In total, the team described about 320 scenarios in which the assistant responds with editorial cues. For each of these scenarios, there are from two to seven answers that the system chooses. randomly... “This is enough to form a personality, but not enough to talk about everything in the world,” says Subbotin.
The developers decided that the assistant would be a young girl who is always ready to help the owner of the smartphone, but at the same time does not allow a familiar attitude. She is correct and keeps her distance when they try to offend her, but she is not arrogant and does not boast of her knowledge.
The assistant must also be ironic and - this is important in the context of systems artificial intelligence- self-ironic. “The fact is that our speech synthesis algorithm - and all the algorithms of our competitors - are not very good at emotion. "Alice" laughs at herself a little, adding the words "joke" or "sarcasm" where necessary. Because she knows about herself that she can’t intonate well yet, ”the developers explain.
How the name was chosen
The team compiled a list of names that employees felt best suited their chosen personality traits. The name was also presented technical requirements: for example, it should not have had the letter "r" - it is not pronounced by small children.
In addition, the name should not have been included in other words or phrases. As one of the options, the developers considered the name Maya - but it could provoke recognition errors on phrases like "May Ninth". It was also important that the name was not included in the list of the most popular female names in Russia - in order to reduce the number of false activations. Alice is not included in such a list, but for those users who have Alice in their family, there will remain the opportunity to activate the assistant with the command "Listen, Yandex".
Having decided on the options, the developers, together with sociologists and analysts from Yandex, compiled a survey for users of the Toloka service. In it, the company asked to determine the character traits of the girl by name. “Users cannot be asked directly which name is more appropriate for an assistant, and we asked slyly. They took the characteristics we needed - and vice versa, completely inappropriate, and began to ask: if a girl has such a name, is she more likely, for example, kind or aggressive? "
By a wide margin, says Subbotin, the name "Alice" won the poll. "And no one ever told us it was a bad name during our testing."
How "Alice" learned during testing
In total, several tens of thousands of people took part in testing the assistant, and it lasted five months.
The main thing that we have found out during this time is that we are moving in the right direction. And that the questions that we thought the users should be concerned about really interest them.
Before starting testing, it was important to determine the most likely scenarios of interaction: the questions that users will ask "Alice" - and also understand how they will ask them.
“We understood that people care about, for example, the weather. Great, to tell about the weather, you need to react to something - to some custom phrases, ”says Subbotin. Having decided on the topic, the developers described the simplest scenarios: for example, when the user simply asks: "What is the weather like now?" Then added more complex options: forecast weather for a specific date, for several days, for several days in the next month.
Another popular topic is location determination. Initially, the Alice team collected several replicas that users could use to activate the script, but it turned out that often people ask questions in a completely different way than the developers might have expected. “We included: 'where am I', 'where am I', 'what is my location', 'what is my address now'. And the user asks: "Tell me, please, where is my geolocation." We could not have foreseen this. "
Developers keep track of the types of user requests and the relevance of the assistant's responses, and add new scripts to the assistant that were not originally built into the assistant.
The team is also developing a special dialogue model - within the company it is called a "chatterbox" - which in the future can take over all communication with the user on abstract topics and learn new cues on its own.
Shoutbox learns on large arrays of texts, including dialogues. Sometimes, the creators say, they cannot predict the answers of "Alice" - they turn out to be so human - and sometimes she can be nasty to the interlocutor. But the developers track the incorrect replicas and remove them from the system.
Yandex has launched Siri analog – voice assistant"Alice". We did not hesitate to give "Alice" a strength test.
According to Yandex, the uniqueness of Alice lies in the fact that she can not only respond with a prepared set of phrases, but also improvise. You can talk to "Alice" without trying to formulate your request as clearly as possible, but with a simple human language, and she will understand you. "Alice" has a character, she can be sad or happy depending on the situation, she has a sense of humor, and if necessary, she can even be insolent. We downloaded "Alice" and checked how true it all is.
When you hear "Alice" for the first time, you feel that you have known her for a long time. The fact is that the voice assistant speaks in the voice of Scarlett Johansson, or rather the actress Tatyana Shitova, who dubs Hollywood star in all films. Including in the film "She", where Johansson plays just the voice assistant, with whom the main character falls in love.
The acquaintance begins with "Alice" talking about her capabilities. In fact, it does everything the same as Siri - it looks at the weather, builds routes, looks for necessary information in the Internet. At the same time, we are told that "Alice" will understand the question, regardless of the wording. Okay, let's do a little check. Suppose I want to know if it will rain in the evening, but I will not ask directly, but as in life: "Should I take an umbrella to work today?" Unfortunately, “Alice” did not understand such a formulation of the question.
I ask "Alice" how long I need to go to work - the assistant gives two options for an answer, depending on whether I go by car or by public transport... Everything is good here. Here are just the next question "Where can I go to have a bite on the way" for some reason, the program gives the address of an esoteric store. Of course, there may be a buffet or a cafe at the store, but this is clearly not the best snack option in Moscow.
"Alisa" immediately answers me who sings a song about "Louboutins", but when asked when "Leningrad" has the next concert in Moscow, she gives me the address of one of the points of theatrical and concert box office called "Concert". In general, the promised technology of communication "as in life" and understanding of the daily manner of speaking is still poor.
During beta testing this spring, many users paid attention to the fact that "Alice" does not even know such simple things as adding two and two, and for an answer turns to the search for "Yandex". These problems have been eliminated in the current version: "Alice" has learned to count and perform other mathematical operations, for example, convert dollars to rubles and vice versa.
Nevertheless, as soon as "Alice" does not understand what you mean, which happens quite often, she runs to look for an answer in a search engine. Often it turns out stupid, as, for example, it came out with my request to recommend a funny movie.
Why is Alice better than Siri? We decided to ask the voice assistant herself, but we never got a clear answer.
Yes, with an understanding of the questions, "Alisa" is not all perfect yet, but she has one more interesting feature- the ability to improvise, thanks to which you can just communicate with her when bored. "Alice" really knows how to joke, both to order and of her own free will. True, apparently the stock of jokes is small, since "Alice" is often repeated.
As for the intellectual development of Alisa, Yandex took care of this very badly. Voice assistant, according to her own assurances, adores the classics, but does not see the difference between Dostoevsky and Tolstoy.
A couple of weeks ago, "Alice" was appreciated by Russian President Vladimir Putin, but even with him, the program did not behave as it should. To the question "Are they not offending you here?" Alice for some reason answered "Okay, I'll take it into account." After such an embarrassment in front of the first person of the state, I would like to advise "Alice" only one thing: study, study and study again.
In October 2017, Yandex announced the release of a new voice assistant Alice. This application is a program that is able to solve everyday tasks of the user, such as: announce the weather forecast, open files on the computer using voice commands, search for information on the Internet and much more. Distinctive feature assistant Alice is that she communicates not using template templates, as most voice assistants do, but with the help of comprehensible human speech, forming responses using neural networks. This allows Alice to answer competently, logically, and sometimes very fun. After the assistant was released to the masses, the application began to be tested and in every possible way to knock out the usual merry people from the rut of the usual conversation. So jokes in Alice Yandex appeared on the Internet, which will be discussed later.
When the conversation looks like a normal dialogue with a real assistant, you ask her a question - she immediately answers it. You can still use the keyboard and mouse for this, if you are using. You can criticize or reward the responses using the ratings in the app.
To dub Alice, Tatyana Shitova was invited, who voices all the roles of Scarlett Johansson in the Russian film distribution. The voice in the application turned out to be very nice and pleasant. After talking with Alice for a while, you can see that she has her own original character. For example, in some of her incorrect answers to a question, she claims the opposite and continues to prove her case for some time.
Alice's sense of humor from Yandex and jokes
When communicating with an assistant, you can catch her original sense of humor and gags. She has personal preferences in many areas, for example, her favorite performers: oxxxymiron, Mushrooms, Dima Bilan and others. Having asked her this question: “Alice, what is your favorite song?”, She replies: “The ice is melting between us, let no one find us now. We'll get wet in the rain, and today we are just the two of us. Tyts-tyts. " Funny, isn't it?
Usually Alice answers one question posed - with one answer. She can also be asked to sing, the reaction in almost all cases is different. She can even "show off" as a real female representative. But more often, he nevertheless agrees without bickering, "Alice, sing a song" - "We were standing on a plane ... It seems that she took the wrong key. Let me rehearse with myself. " She gives out such jokes quite often, which cannot but cheer up even in bad weather.
Sometimes our friend suddenly takes offense at us if, for example, we do not allow Yandex to determine our location and says that we are not telling her anything. If you start to be indignant, Alice also perceives this as an insult and hints that we are poorly educated people.
If you suddenly get bored, Alice can be asked to tell an anecdote or a fairy tale. She does it well. When you ask her a question about what she can do, sometimes she replies that she knows how to give everyday advice of dubious value. The programmers did a good job developing Alice's applications.
Disadvantages and features of the voice assistant Alice
There are also gaps in the work of the assistant or. For example, she is not always able to open the application that you dictate to her. On some questions, instead of answering by voice, she opens links in the browser with the specified request... Alice examines great amount text on the Internet, but still does not cope with some tasks. The neural network allows you to study both positive texts with a pleasant manner of communication, and negative statements that are found on the network. To the request "Order me a sandwich", Alice can answer you - "You will manage". In this answer, you can see both joke and humor, as well as sharp sarcasm. Even so, like a child, she can be taught good conversation. If you don’t tell her that she’s talking in the wrong way, you may hear something different from her in the future.
Possible development of Alice
The development of the voice assistant Alice began at the end of 2016. The industry is moving in this direction, and the user is ready to communicate with the computer assistant in normal human language and indicate certain actions through dialogue. With the help of the Alice application, this approach is fully implemented.
At one of the speeches, a Yandex representative said: “If people have the opportunity to speak in a normal language with a computer, why not implement it in the Alice application. If a person needs to ask a search engine a question - where to buy an engine for a Lada Priora sedan - then let him have the opportunity to do it in natural language, and not first formulate the entire query, and only then ask it to the search engine.
In contact with
