Dasha conversational AI as a Service. What are the ML services giving your apps human-like magic?

Dasha voice Artificial Intelligence as a Service stack overview

Take a look at the architecture map. The Voice Processing, NLP and Conversation Model areas in the application layer are what we will be talking about today. For the purpose of today’s discussion, we are talking about conversations executed with voice, not text, even though Dasha AI does text conversations as well.

Voice processing

Let’s start at the end, with voice activity detection. There is a good reason we’re starting with this configuration, as it detects speech activity on the user’s side. This is used to signal to Dasha application to listen or to check in if the user is still there. I won’t talk about every property used in this config (you can find details in our documentation but will mention interlocutorPauseDelay because I think it’s a great tool. It lets you set how many seconds’ (to a decimal) pause should Dasha take before responding to a human’s reply. Just like in a conversation you may have with a live person, you don’t want to barge right in with a response but may want to let a person collect their thoughts and finish their statement. This makes for a more productive conversation and you can teach Dasha to behave in such a polite manner with a simple parameter.

On to speech to text (STT) which is responsible for translating human speech into machine readable text. You can use the default speech recognition service, courtesy of Dasha Platform

Text to speech (TTS) - this is where things get interesting. You can connect your own third party TTS service. The vast majority (if not all) or our users use Dasha default TTS. Our synthetic speech is our pride and joy. Almost half of our machine learning engineers work on constantly improving our speech synthesis. We have a pretty simple goal - to run a conversational speech engine that passes the Turing test every time.

As it were now, we roll out updates once or twice a month and there are a number of configuration properties that you can use to modify the speaker’s voice for the purposes of your Dasha conversational AI application.

NLP - natural language processing

text classification and named entity recognition are two key parts of Dasha’s natural language understanding engine. NLU ensures that the neural brains understand the human’s speech with the meaning that the human has put into their words.

Starting from the left - text classification is often referred to as intent classification. There are neural networks operating here that extract specific meaning (intent) from the speaker’s words. This is an absolutely key functionality that you can control by modifying the intents.json file. In the nutshell - you provide a set of training data points (5-20 phrases) for a specific intent. The neural network then uses the data to train itself to fully recognize the intent in the scope of a conversation. Ilya (ML team lead) wrote a detailed post on how you can use intent classification for a robust conversational AI app.

Named entity recognition (NER) is a super fun concept. Let’s say you want your AI app to take a food order. You need to let it know that at a specific point in the conversation the user will say a number of phrases which will signify a specific set of food items. You would use NER for that. By the same token, if you wanted to take the user’s name, you would use NER to teach the app at which point in the conversation to expect to hear the name. Here is Ilya again with a write up on how to use Named Entity Recognition aspect of this AI as a Service startup to make your app communicable and personable. And for a tutorial using NER, check out this one on building a food ordering app from yours truly.

Sentiment analysis mentioned in the map refers to a feature that classifies the user’s sentiment as “positive” or “negative”. It was used to classify yes/no responses. This a feature is in need of an update, so I will suggest you not use it and instead create a custom intent for “yes”/”no”, as described in the tutorial above.

Natural language generation refers to the ability of the speech synthesis (TTS) engine to generate human-like speech. Each Dasha project has a phrasemap.json file. In the file you can specify such parameters as variations in voice:

voiceInfo: {
    speaker: string
    lang: string
    emotion?: string
    speed?: number
    variation?: number
  }

Or construct whole phrases with variables to be used in the course of the dialogue. (Pro tip: if you’re a pro, you may want to do this. But first, play around with using phrases directly in the DashaScript (main.dsl). The thing is, as you’re speaking to your buddy, he may ask you to repeat a phrase and you will paraphrase or shorten your speech. You can do the same with Dasha’s NLG service in the AI cloud platform:

interface IRandom<T> {
  random: T[]
}
interface IRepeatable<T> {
  first: T
  repeat?: T
  short?: T
}

Personally, I love playing with the phrasemap. Here is a longer description in our documentation.

Conversation model

Arguably, this is the part that makes conversational AI apps built with Dasha unique in their human-likeness. Here is a whole set of goodies that make the end user feel listened to and understood.

I’ve started at the end and in the beginning. Let’s go to the middle now and look at digressions. When talking to your mom about your latest pet project she may suddenly ask “say, remember that boy Jimmy you were friends with in seventh grade? I wonder what he’s been up to.” That, my friend, is a digression. A curveball flying out of the blue to catch you off guard. The thing is that you are a human and your brain is able to adjust, go off on the conversational tangent and then come back to the heart of the discussion you were having with mom.

Digressions let you prepare your Dasha conversational app for tangential conversations that people may throw at it. Digressions can be called up at any point in the conversation by the user. Now, the Jimmy question is truly a curve ball and it’s doubtful that you could have planned for it. However, you can expect mom to ask a variety of digressing questions such as, “how is the weather,” “have you seen the latest episode of x,” “are you eating veggies,” “have you been outside this week,” or “do you need help with laundry” and you can prepare for these questions by doing two things. A) create a corresponding intent in intents.json B) create a digression that is activated when the intent is recognized in your main.dsl DashaScript file.

See here for more on digressions. And no, I am by no means encouraging you to automate your conversation with mom.

Interruptions are just that - a parameter that lets you define one of the AI’s phrases as either interruptible or not.

Common sense logic comes as part of the common libraries with every Dasha app.

Turn taking, context reasoning, dictation mode are all implemented within DashaScript code. Slot filling is solved by the NER engine.

And there you have it. Now you’ve got an overview of the AI as a Service capabilities served by Dasha Platform to make your AI apps more human-like and conversational than ever before.

To get started building, join our developer community and get your API key.