Guest Blog - All About Intents - Mycroft Intents

Originally published at: Guest Blog - All About Intents - Mycroft Intents - Mycroft

Originally posted at

So you’re interested in making voice skills but don’t know where to begin? There’s a lot of jargon being thrown around with Adapt, Padatious, context, converse, etc… It’s easy to get intimidated at first, but have no fear because this primer is going to help get you started!

In my day job, I spend my time building an educational system called Chatterbox ( that teaches kids skill design. I’ve realized that teaching kids is much like teaching a novice. I’m challenged to clearly express what intents and utterances are, how they are meant to be used, as well as how to navigate the constraints of a voice-first experience.

Although we all inherently understand what intents are, for a lot of people and especially kids, these concepts and words aren’t used in their day to day life.

If I simply say “intent” what pops up in your mind?

IF… THEN… statements?

You just gave Mycroft an order, how does it decide what to do?

It helps to think about it in terms of if… then… statements

if user said "hello" then say "hello"
if user said "thank you" then say "you are welcome"

When starting a skill, the first thing you should do is think of these if… then.. statements.

You don’t need to write any of this down, but it helps to have a list with the bare minimum of intents you need for the functionality you are trying to implement. Do not worry about how the user will ask the question yet, just keep one statement for each possible outcome.

What is an intent for Mycroft? Each possible outcome of the conversation you are modeling is an intent, that’s why I suggested you write those if… then… statements. Each of those is an intent you need to teach Mycroft!


Do we have to code these if… then… statements? Heck no, we are not in the 80’s, there’s real AI now!

The simplest way to get started is to use Padatious. With Padatious, you just give it examples and it knows what to do.

You do not need to understand the next paragraphs at all to use Padatious, but if you are curious of the inner workings, read on.

Padatious is Mycroft’s neural network intent parser, it will ensure exact matches to any example phrase you give it by using its old fashioned regex cousin Padaos. But thanks to the AI, it also manages to understand sentences you did not explicitly teach it.

Each Padatious intent is a neural network model that tells Mycroft if it should select a specific intent given a request, with a probability. The Padatious engine chooses the highest scoring intent, or none if all probabilities are below a threshold.

Specifically, it uses one-hot encoding via local vocabularies (and a few extras like an unknown ratio input) into a series of shallow feedforward networks to perform intent recognition and entity extraction.

If you didn’t understand any of that, it’s ok; using it in practice is super easy! There is a command line tool named msk (the Mycroft Skills Kit) that creates the skill template for you. You can even create complete skills if they are really simple. You can learn more about msk here.

Let’s make a real skill, LearnPortugueseSkill.

if user said "say hello in portuguese" then say "olá"

How many ways can the user give that request?

Either use msk or create the files manually


The file should contain all sample sentences you can think of for this intent

say hello in portuguese
how do you say hello in portuguese
what is the portuguese word for hello
translate hello to portuguese

Now let’s look at our skill code, open the __ init file

from mycroft.skills.core import MycroftSkill, intent_file_handler

class LearnPortugueseSkill(MycroftSkill):

def handle_hello(self, message):

def create_skill():
return LearnPortugueseSkill()

Since this is a teach Portuguese skill we need to implement some logic,

Mycroft’s TTS system is configured for English for now; it can not speak Portuguese text. Thankfully, the awesome Translate Skill from @jcasoft uses google TTS for this.

Let’s use his code to add speak_portuguese(text) and translate_to_portuguese(text) methods to our skill!

We will now be able to speak the Portuguese bits

from mycroft.skills.core import MycroftSkill, intent_file_handler
from import wait_while_speaking, play_mp3
import os

class LearnPortugueseSkill(MycroftSkill):
path_translated_file = “/tmp/portuguese.mp3”

def translate_to_portuguese(self, text):
    translated = translate(text, "pt")

def speak_portuguese(self, sentence):
    get_sentence = 'wget -q -U Mozilla -O ' + self.path_translated_file + \
                   '"' + \
                   str(sentence) + '&client=tw-ob' + '"'


def handle_hello(self, message):
    # wait=True makes mycroft wait until speech is finished playing before continuing
    self.speak_dialog("hello_in_portuguese", wait=True)

notice that self.speak_dialog("hello_in_portuguese") ? That line of code is using a dialog file, msk or you should have created it at


The file should contain all sample sentences Mycroft can speak to the user, and one will be picked randomly.

hello in portuguese is 
you say it like this

To add support for new languages, just create a new folder with its language code. You can edit your examples at any time without touching the code, the skill will even auto reload. Simple, isn’t it?

It’s also easier to translate an intent if you have full sentences. Adapt is trickier and often requires the translator to check the source code. That’s just one more good reason to use Padatious!

NOTE: Mycroft recently changed the folder structure, all files are now in skill_folder/locale/en-us. Old paths still work, but now it’s easier since you need to track even fewer folders!

I like Rules

if user said "say thank you in portuguese" then say "obrigado" and explain gender

If you are like me, it is easier for you to think of rules than to come up with examples, or maybe you just need more control over when the intent triggers.

Adapt comes to the rescue here; instead of magically learning from examples, you can specify some rules for when you want it to trigger.

def handle_thank_you(self, message):

For this you need to create a file with keyword examples

# ThankYou.voc
thank you


in portuguese
in portugal
in pt
in p t

Extracting Keywords

Every keyword you defined for Adapt can be used as a variable in the intent code. Let’s improve our Adapt intent to use a gender keyword.

def handle_thank_you(self, message):
    gender ="gender")
    if not gender:
        self.speak_dialog("if_male", wait=True)
    elif gender == "male":
    elif gender == "female":

This is the main use case for optional keywords, but they can also be used to increase confidence and help in disambiguation. After having the bare minimum requirements feel free to add as many optional keywords as you want.

def handle_thank_you(self, message):

But what if we want to extract something from the text that we don’t have an example of?
Padatious makes it super simple, in your .intent file write your examples like this

say {{words}} in portuguese
  speak {{words}} in portuguese
  translate {{words}} to portuguese

That’s it, you can now access it just like with Adapt

  def handle_say(self, message):
    words =["words"]

What about multi-turn dialog?

I have written a blog post about this before, so I don’t want to repeat myself a lot.

Adapt intent parser supports contexts. A context is a keyword that becomes available even if the user didn’t say it.

This can be used for continuous dialog, so you can provide data with it (the missing keyword value).

if user said "say again" then repeat last portuguese sentence

Let’s implement repeat functionality, first create repeat.voc

say that again

Now let’s change the speak_portuguese method to set a context

def speak_portuguese(self, sentence):
    get_sentence = 'wget -q -U Mozilla -O ' + self.path_translated_file + \
                   '"' + \
                   str(sentence) + '&client=tw-ob' + '"'
self.set_context("previous_speech", sentence)

The context made the “previous_speech” keyword available to Adapt. This intent can now be triggered up to 3 questions after translate_to_portuguese was last triggered.

def handle_repeat(self, message):
    text ="previous_speech")

Padatious does not yet support context; for these cases, you are stuck with Adapt. However, you can set and remove contexts at will inside Padatious intents like I just did in the speak_portuguese method.

You can also use contexts that don’t even have data, you just require it to ensure that something happened first

def speak_portuguese(self, sentence):
    get_sentence = 'wget -q -U Mozilla -O ' + self.path_translated_file + \
                   '"' + \
                   str(sentence) + '&client=tw-ob' + '"'
self.set_context("previous_speech", sentence)

def handle_how_do_you_know(self, message):

I am a wizard

So I hear you like regex and Padatious is not an option for you, or maybe you are afraid because you don’t understand how Padatious learns, but you aren’t the botphobic kind are you?

if user said "explain {SampleWord} in portuguese" then say dictionary definition in portuguese

For this I will be using PyDictionary. You can add it to your skill by creating a requirements.txt file and adding PyDictionary.
I always recommend you use Padatious, but sometimes regex is a necessary evil and Pythex is useful to check your regex.

Make a file named dictionary.rx

explain (?P.*) in portuguese
  meaning of (?P.*) in portuguese

Just require and use the regex capture group as a normal Adapt keyword.

def handle_explain(self, message):
    word =["SampleWord"]
    dictionary = PyDictionary()
    meaning = dictionary.get("Noun") or dictionary.get("Verb")

Padatious is much more human-readable, easier to translate, and less prone to errors. Adapt’s regex is also known to be somewhat buggy at times, but maybe you really are a wizard.
Just keep in mind Adapt is rules-based; corner cases creep up if you oversimplify or overcomplicate.

"hey mycroft, where is the nearest bus stop"
"stopping everything"

"hey mycroft, do i need to say the last order again?
“the last order again”

I am not a wizard!

Ok, you are still having trouble, there’s some corner case you can’t figure, need to extract a date or a number maybe?

if user said "say {weekday's date} in portuguese" then translate date to portuguese
if user said "say {number} in portuguese" then translate number to portuguese

We have utility packages to extract dates, and English is well supported

def handle_date(self, message):
    date, text_remainder = extract_datetime(["utterance"], lang=self.lang)
    pronounced_date = nice_date(date)

One useful strategy that works well with optional keywords is to use the utterance_remainder.

In Adapt intents you can get the text leftover that was not captured into any keyword.

Mycroft also provides utils to handle numbers, language support not guaranteed except for English.

The PR for Portuguese is in, we can pronounce the number directly and save a call to google translate

def handle_number(self, message):
    text = message.utterance_remainder()
    # lets get a number from the utterance
    number = extract_number(text, lang=self.lang)
    # portuguese uses long scale, lets take that into account!
    # in long scale 1 billion = 1e12 instead of 1e9
    spoken_number = pronounce_number(number, lang="pt", short_scale=False)

No intents needed

Need an answer to a general purpose question? Mycroft also has a mechanism to try to answer things it doesn’t have intents for.

Fallback skills are tried by order until one can answer you, this is a good place to plug skills like Wolfram|Alpha.

But sometimes you just could not make an intent for your use case, maybe because there are a lot of skills and there are intent collisions.

if user said "say a pun in portuguese" then say a pun only portuguese speakers can understand

A last-resort thing you can do is make a fallback intent using good old fashioned python programming with no help whatsoever to decide what to do.

class LearnPortugueseSkill(FallbackSkill):
    puns = [...]
def initialize(self):
    self.register_fallback(self.handle_pun_fallback, 99)

There’s usually no good reason to do this, and this example should have used a regular intent.

def handle_pun_fallback(self, utterance):
    if self.voc_match(self, utterance, "pt_pun"):
        pun = random.choice(self.puns)
        question = pun["pergunta"]
        answer = pun["resposta"]
        self.speak_portuguese(question + ".\n" + answer)
        return True
    return False

Intercepting the intent parser

What if we want to capture the whole utterance cycle inside a skill without giving other skills a chance?

if user said "translate everything i say to portuguese" then start ignoring questions and translate everything

Let’s make an intent to start the interception cycle

def handle_live_translate(self, message):
    self.speak_dialog("start_tx", wait=True)
    self.speak_portuguese("iniciando tradução automática")
    self.intercepting = True

Now we can use the converse method to capture the utterances.

But we also need to check if the user told us to stop. There is a helper method to check if some .voc file would match.

def stop(self):
    if self.intercepting:
        self.speak_dialog("stop_tx", wait=True)
        self.speak_portuguese("parando tradução automática")
        self.intercepting = False

def converse(self, transcript, lang=“en-us”):
utterance = transcript[0]
if self.intercepting:
if self.voc_match(self, utterance, “cancel”, lang=lang):
return True
return False

Now every time you say “Hey Mycroft” it will repeat what you said in Portuguese and ignore any questions.

So what did we learn?

I wanted to show you the potential of Mycroft intent system. Hopefully Mycroft is less confusing for you now!

Some functionality of this example skill could probably be made other (better) ways, but it should have taught you to:

– speak pre-programmed answers, in our case translate hello to Portuguese
– optionally take extra parameters into account, gender matters when speaking in many languages
– extract keywords with Padatious and translate them to Portuguese
– extract keywords with regex and explain their meaning in Portuguese
– pass data to follow up intents (repeat last Portuguese utterance)
– use Mycroft helper utils, utterance_remainder, extract/pronounce number and extract/pronounce date
– create a fallback handler and use custom utterance parsing
– take control of the utterance processing cycle using converse method (live translate anything to Portuguese)

Source code is available here, and your skill folder should look like this

├── locale
│   └── en-us
│       ├── date.voc
│       ├── dictionary.rx
│       ├── gender.voc
│       ├── google_told_me.dialog
│       ├── hello_in_portuguese.dialog
│       ├── hello_in_portuguese.intent
│       ├── if_female.dialog
│       ├── if_male.dialog
│       ├── inPortuguese.voc
│       ├── know.voc
│       ├── number.voc
│       ├── pt_pun.voc
│       ├── question.voc
│       ├── repeat.voc
│       ├── say.intent
│       ├── say.voc
│       ├── start_tx.dialog
│       ├── stop_tx.dialog
│       └── ThankYou.voc
└── requirements.txt

Now if you wanted to show this skill to the world, all you need to do is create skill tests and submit the skill to the Marketplace!

Only for Mycroft?

Adapt and Padatious are open source, so you can use them outside Mycroft in your own projects! In fact, Mozilla is using Adapt in an IoT project.

Installing Padatious is easy

sudo apt-get install libfann-dev python3-dev python3-pip swig
pip install padatious

Installing Adapt is also easy

pip install adapt-parser

Padatious is really simple to use

from padatious import IntentContainer

container = IntentContainer(‘intent_cache’)
container.add_intent(‘hello’, [‘Hi there!’, ‘Hello.’])
container.add_intent(‘goodbye’, [‘See you!’, ‘Goodbye!’])
container.add_intent(‘search’, [‘Search for {query} (using|on) {engine}.’])

print(container.calc_intent(‘Hello there!’))
print(container.calc_intent(‘Search for cats on CatTube.’))


You can also get Padaos the regex intent parser. I even forked it to simply output the generated regex instead of making intents.

from padaos import IntentContainer

container = IntentContainer()
container.add_intent(‘hello’, [
‘hello’, ‘hi’, ‘how are you’, “what’s up”
container.add_intent(‘buy’, [
‘buy {item}’, ‘purchase {item}’, ‘get {item}’, ‘get {item} for me’
container.add_intent(‘search’, [
‘search for {query} on {engine}’, ‘using {engine} (search|look) for {query}’,
‘find {query} (with|using) {engine}’
container.add_entity(‘engine’, [‘abc’, ‘xyz’])
print(container.calc_intent(‘find cats using xyz’))

{‘name’: ‘search’, ‘entities’: {‘query’: ‘cats’, ‘engine’: ‘xyz’}}

Adapt seems a little more intimidating at first sight.

from adapt.intent import IntentBuilder
from adapt.engine import IntentDeterminationEngine

engine = IntentDeterminationEngine()

weather_keyword = [“weather”]

for wk in weather_keyword:
engine.register_entity(wk, “WeatherKeyword”)

weather_types = [“snow”, “rain”, “wind”, “sleet”, “sun”]

for wt in weather_types:
engine.register_entity(wt, “WeatherType”)

weather_intent = IntentBuilder(“WeatherIntent”)


for intent in engine.determine_intent(’ '.join(sys.argv[1:])):
if intent.get(‘confidence’) > 0:
print(json.dumps(intent, indent=4))

Jarbas is a long time Mycroft Community Member and a developer at Chatterbox, a company teaching kids about technology through a build-it-yourself, code-it-yourself smart speaker.

Check out Chatterbox on Kickstarter!
Read more tutorials from Jarbas


I do not know if it just me or what but I can not get these to install or work on newest armbian (sketch) or Ubuntu flavour lubuntu 19.04 ( dropped my big desktop because of its >200watt consumption in favour of a small micro computer alfawise z83 for its <20watt consumption)

msk installs but no way to make it run that I can see
and Padatious just complains fann lib is not installed even though they are

Hey there,
Are you installing these independently or bundled together with Mycroft-core?

If you installed mycroft-core, what happens when you attempt to run mycroft-msk create?
If you installed MSK independently this would be just msk create

I’m not familiar with alfawise’s hardware but looking at the specs there’s nothing that immediately tells me it couldn’t work. There are however always quirks when getting something to run on new hardware.

For the libfann error, what output do you get for:
apt list --installed | grep fann

thank you kindly for the reply

it was with the mycroft-core so `mycroft-msk create worked. --thank you
to bad they did not mention this in the github readme or the wiki would of made it simpler

Padatious - I am not sure it it still required since I am installing with the core. but here is the grep out put

libfann-dev/disco,now 2.2.0+ds-5 amd64 [installed]
libfann2/disco,now 2.2.0+ds-5 amd64 [installed,automatic]

Thanks for flagging that about the MSK docs, I’ve updated our primary documentation.

The libfann packages look correct and Padatious certainly gets installed with Mycroft-core already.

mycroft-msk create throws a bunch of errors:

(.venv) pi@picroft:~ $ mycroft-msk create
Traceback (most recent call last):
File “/home/pi/mycroft-core/.venv/bin/msk”, line 5, in
from msk.main import main
File “/home/pi/mycroft-core/.venv/lib/python3.7/site-packages/msk/”, line 23, in
from msk.actions.create import CreateAction
File “/home/pi/mycroft-core/.venv/lib/python3.7/site-packages/msk/actions/”, line 34, in
from msk.console_action import ConsoleAction
File “/home/pi/mycroft-core/.venv/lib/python3.7/site-packages/msk/”, line 21, in
from msk.global_context import GlobalContext
File “/home/pi/mycroft-core/.venv/lib/python3.7/site-packages/msk/”, line 23, in
from msk.util import ask_for_github_token
File “/home/pi/mycroft-core/.venv/lib/python3.7/site-packages/msk/”, line 25, in
from git.config import GitConfigParser, get_config_path
ImportError: cannot import name ‘get_config_path’ from ‘git.config’ (/home/pi/mycroft-core/.venv/lib/python3.7/site-packages/git/

Try installing GitPython:

pip install --upgrade GitPython