SpeechHandler Plugin Development

Parent Class: plugin.SpeechHandlerPlugin
Required Methods:
intents()
    Parameters:
      none
    Returns an intents structure

handle(intent, mic)
    Parameters:
      intent: a returned intent structure
      mic: allows naomi to listen and speak
    Returns null

Optional Methods:
settings()
    Parameters:
      none
    Returns a settings structure

Writing a speechhandler plugin

So you installed Naomi and had a pleasant conversation about the time and weather. Now you have an idea for something you would like Naomi to do, but aren’t sure how to start working on it. We welcome new ideas for Naomi, and now we have a store where you can make your plugin available to everyone using Naomi. Let's walk through the process of creating a SpeechHandler plugin.

Start by creating a new project on Github or some other online git repository (the Naomi Project does not host the plugins in the store, just links to them, so you need to put your source code someplace where others will be able to access it). Give your project a name and select a license (we use the MIT license, so if you choose a more restrictive license like GPL and decide to distribute Naomi with your plugin, you can distribute the whole thing under GPL. If you would like the NaomiProject to consider bundling your plugin with Naomi, you should use the MIT license).

Now, navigate to the ~/.config/naomi/plugins/speechhandler folder (or wherever your local settings folder is located, this is the default, so this is where it will be unless you changed it) and git clone your repository (if that sentence didn’t make any sense to you, try searching for “git clone” online). Now create the following files in your cloned project folder:

  • plugin.info (basic information about a plugin, like name and version)
  • __init__.py (this lets python know that this is a package)
  • yourproject.py (this can be named anything)

There are a lot of examples of packages in the plugins/ directory under the main Naomi folder, so you can look to them for examples of what can go into this file. Here are the most important parts:

plugin.info File

The info file is set up with major sections marked by a section header which is surrounded by square brackets ([Plugin]) followed by Name = Value pairs. Naomi only looks at the [Plugin] section, and only requires the Name, Version, URL and Description fields. Your plugin must either contain a license file (LICENSE, LICENSE.md, LICENSE.txt) or have a value in the License line under the [Plugin] section. Version and Description are only used when listing or installing plugins. The URL is used to determine whether you are trying to update an existing plugin, or install a new plugin with a similar name.

Here is a typical plugin.info file:

[Plugin]
Name = You are Welcome
Version = 1.0.0
License = MIT
URL = https://github.com/aaronchantrill/yourwelcome.git
Description = Responds to “Thank you” with “You are Welcome”

You are welcome to include additional information and add additional sections, but this is all that is required. The sections that users most commonly add are [Author], [Contact] and/or [Support].

__init__.py

Now the __init__.py file provides the actual entry point into your plugin. In fact, you could put your entire plugin into the __info__.py file, but traditionally people tend to keep this file pretty sparse and just use it to import the main module:

from .youarewelcome import YouAreWelcomePlugin

Main plugin module

So far, pretty simple. Now the main part of the plugin contains a class with a couple of specific methods.

from naomi import plugin

class YouAreWelcomePlugin(plugin.SpeechHandlerPlugin):

    def intents(self):
        return {
            'Youre_WelcomeIntent': {
                'locale': {
                    'en-US': {
                        'templates': [
                            self.gettext('Thank you'),
                            self.gettext('Thanks')
                        ]
                    }
                },
                'action': self.handle
            }
        }

    def handle(self, intent, mic):
        response = random.choice([
            self.gettext("You are welcome"),
            self.gettext("You’re welcome"),
            self.gettext("Don’t mention it"),
            self.gettext("It was nothing, I’m serious, it meant nothing to me")
        ])
        mic.say(response)

intents

The first method that every SpeechHandler plugin has to have is “intents”. This returns a dictionary structure telling Naomi how you expect people to activate your plugin. Here is a typical intents method:

def intents(self):
    return {
        'YouAreWelcomeIntent': {
            'locale': {
                'en-US': {
                    'templates': [
                        "Thank you",
                        "Thanks"
                    ]
                }
            },
            'action': self.handle
        }
    }
}

This is a simple intent with only templates for a single locale and action. This is the minimum required. The templates node contains a single list of things someone might say to activate this plugin, and the action node contains a reference to the entry method for the plugin. Note that this is an actual reference to the method, not just the name of the method, and the method does not have parenthesis after it because you are passing a reference to the method itself, not the result of calling the method.

There are currently two other elements that can be added to an intent, keywords and regex.

keywords are a list of words that can be matched by the intent parser. For instance, Naomi's confidence would most likely increase if you rewrote the above intent as:

def intents(self):
    return {
        'YouAreWelcomeIntent': {
            'locale': {
                'en-US': {
                    'keywords': {
                        'ThanksKeyword': [
                            'thank you',
                            'thanks'
                        ]
                    },
                    'templates': [
                        '{ThanksKeyword}'
                    ]
                }
            },
            ‘action’: self.handle
        }
    }

This way, the intent parser will see '{ThanksKeyword}' as a single word that appears in every template for this intent, and not in any of the templates belonging to other intents.

regex are currently only used by the Adapt intent parser, but they are so central to the way Adapt and other atomic type intent parsers work that we had to include it. This allows you to extract a value from a phrase based on a regular expression. For instance, if you wanted to match and return the value to be searched for from the requests "search for cats on youtube" or "search youtube for cats", you could write an intents method like this:

def intents(self):
  return {
    'SearchIntent': {
      'locale': {
        'en-US': {
          'keywords': {
            'EngineKeyword': [
              "google",
              "youtube",
              "instagram"
            ]
          },
          'regex': {
            'Query': [
              " for (?P<Query>) on ",
              " for (?P<Query>) using ",
              " for (?P<Query>.*)$"
            ]
          },
          'templates': [
            'search for {Query} using {EngineKeyword}',
            'search for {Query} on {EngineKeyword}',
            'search {EngineKeyword} for {Query}'
          ],
        }
      },
      'action': self.handle
    }
  }

handle

The handle method must have the following declaration:

def handle(self, intent, mic):

The structure passed to the "intent" parameter tells the plugin why it was activated and may contain some information about matched words, and the object passed to the “mic” parameter can be used to communicate with the user. For example, the SearchIntent above would return a return intent structure like this:

{
  'action': &lt;function &lt;lambda&gt; at 0x7f85d000b6a8&gt;,
  'input': 'SEARCH FOR CATS ON YOUTUBE',
  'intent': 'SearchIntent',
  'matches': {
    'Query': ['cats'],
    'EngineKeyword': ['youtube']
  },
  'score': 0.8333333333333334
}

From this, it is easy to see that the intent parser believes that the user probably wants to perform a youtube search for cats.

In our YouAreWelcome plugin example, the handle method is really simple:

def handle(self, intent, mic):
  mic.say(self.gettext(“You are welcome”))
  return True

Internationalization

Translating intents into different languages is fairly straightforward, since the locales section allows you to create a translation for a particular locale directly. This allows different locales to define different types and numbers of keywords and templates as necessary.

You might be wondering what self.gettext() is in the above examples. Self.gettext is created in the plugin.GenericPlugin base class that all plugins are derived from, and is used for internationalization. Naomi is an international application and we want to make it available in as many languages as possible. Right now we have good support for English, French and German. If you are interested in helping to add a new language through translations, testing, or documentation, please reach out to the Naomi Project team.

When you run ./update_translations.py, every python program will be searched for gettext(“.?”) and the contents of “.?” will be added to a .pot file. That .pot file will be combined with the current .po file (if there is one) creating a .po file. This file is basically just a list of all of these static strings, with space to add a translation below. For example, here is a small .po file:

# #-#-#-#-#  fr-FR.po (Naomi 2.2-dev)  #-#-#-#-#
# Naomi YouAreWelcomePlugin
# Copyright 2019
# Distributed under the MIT license
#
#, fuzzy
msgid ""
msgstr ""
"Project-Id-Version: Naomi 2.2-dev\n"
"Report-Msgid-Bugs-To: https://github.com/NaomiProject/Naomi/issues"
"POT-Creation-Date: 2019-10-09 19:00+0100\n"
"PO-Revision-Date: 2019-10-10 00:16+0100\n"
"Last-Translator: Aaron Chantrill <aaron.chantrill@dottywood.org>\n"
"Language: fr_Fr\n"
"Plural-Forms: nplurals=2; plural=(n>1);"
"MIME-Version: 1.0\n"
"Content-Type: text/plain; charset=UTF-8\n"
"Content-Transfer-Encoding: 8bit\n"
"Generated-By: pygettext.py 1.5\n"

#: youarewelcome.py:19
msgid "Thank you"
msgstr "Je vous remercie"

#: youarewelcome.py:20
msgid "Thanks"
msgstr "Merci"

#: youarewelcome.py:27
msgid "You are welcome"
msgstr "Je vous en prie"

It is important to note that when the phrases are extracted, the program is not running and none of the variable have values. So if you put your gettext method around a variable, only the variable name is returned to the .pot file. So, for instance, if you do something like this:

response = random.choice([
  “You are welcome”,
  “You’re welcome”,
  “Don’t mention it”
])
mic.say(self.gettext(response))

Nothing will be returned in the .pot file for response. So always do this:

response = random.choice([
  self.gettext(“You are welcome”),
  self.gettext(“You’re welcome”),
  self.gettext(“Don’t mention it”)
])
mic.say(response)

A common mistake is to try to split up text like this:

self.gettext("It was nothing," +
  "I’m serious, " +
  "it meant nothing to me"
)

or this:

self.gettext(" ".join(
    [
        "It was nothing,",
        "I’m serious,",
        "it meant nothing to me"
    ]
))

When you go to create a translation file, you will get an error message like: *** youarewelcome.py:31: Seen unexpected token "+" or *** youarewelcome.py:31: Seen unexpected token "."

You pretty much have to keep the entire line of text on one line between the quotation marks. Also, please don’t enter any punctuation into the intents, other than apostrophes that are part of words. We do not currently clean that input, and punctuation will likely cause an error.

Settings

Now suppose you need some information from the user to set up your plugin. Say you are writing a plugin that uses a web service to list upcoming movie times. You don’t want to list all movie times all over the country, and you don’t want the user to have to tell Naomi exactly where they are every time they want movie information. So when your plugin loads for the first time, the user is asked to enter their zip code so that information can be passed to the web service you are using.

You can add the settings method to Naomi:

def settings(self):
        _ = self.gettext
  return OrderedDict(
            [
                ('zipcode',): {
                    "title": _("What is your zip code?"),
                    "description": _("Providing your zip code will help the Movietime plugin bring you the most relevant movie information")
                }
            ]
        )

The structure returned by the settings method has a number of different fields that can be used to describe how you want to collect information. Right now, this only happens sequentially through the text interface as you start Naomi. We hope to have a web service that will generate forms that can be accessed via a web browser soon. In the meantime, it is important that you start Naomi manually and not as a service after installing a new plugin, in case that plugin needs to request information.

NAME - (required) The name of the variable as seen by the developer. Since this can be a multi-level variable, it will be passed as a tuple containing the nodes to the path. So, to set the value

email:
  address: me@myhostname.net

you would use the tuple ('email','address'). This will form the key for a dictionary containing the whole form.

TYPE - (optional, default "textbox") - This is the general form for the control, without specifics about how to implement it. These fall into the following general categories:

  • A simple text input box (default)
  • “boolean” - A simple yes/no question (only possible responses are yes or no)
  • “combobox” - A combobox where the user can select a value from a list or enter a new value
  • “listbox” - A listbox where the user can only select a value from the provided list (or an empty string)
  • “file” - A file select box where the user can either browse the server filesystem looking for a file, or can upload a file from their local filesystem which will be saved in a location on the server.

Notice that I don't specify radio buttons or checkboxes since these are implementation details covered by the concept of a listbox. We will probably also need a multiple list box where multiple options can be selected, but there isn't an instance where that is currently required for Naomi.

TITLE - (required) A brief (five words or so) description of the setting.

DESCRIPTION - (optional, default "") a longer description of what exactly the setting does.

DEFAULT - (optional, default "") The default value as chosen by the developer of the plugin. This can be a function. This will only show up if the user has not previously selected a value.

OPTIONS - (optional, default []) In the case of a combo or list box, a list containing the available choices. This can be a function.

VALIDATION - (optional, default True) a function used to validate the user response and make sure the user entry is valid. If the value entered by the user is not valid, then the setting will not be saved. On the command line, the question will continue to be put to the user until they enter a valid response. An empty response will always be valid and can be taken to mean that the user has chosen not to respond. Some standard validation functions (for checking things like email address, number, integer, etc.) have been provided for developers.

ACTIVE - (optional, default True) a function used to tell if this should be active or not based on the current state of the form. This is used to disable or enable an option depending on values entered into the system.

Publishing your plugin

So you have made it all the way through and your plugin works just the way you always hoped it would. Congratulations! Now we would love for you to share your creation with us. You did git commit your files and then git push the changes to your online repository, right? If not, please take a moment to do so now.

Now publishing your plugin through Naomi requires forking a new repository and adding a line to a CSV file.

To start, make a fork of the naomi-plugins repository here

Now edit the plugins.csv file and add a line containing information about your plugin. This is very simple, so you can just use the online editor on GitHub. There’s no need to clone the file locally, but you can if you like.

Fill in the fields of the table. Most of this information (Name, Version, Description) can be copied directly from your plugin.info file. The Repository field should be the URL from the plugin’s [Plugin] section, the category should be the type of plugin (audioengine, speechhandler, tti, tts, stt, stt_trainer, vad, or visualizations).

The last field is "commit". This will identify the actual git commit that will be installed by the store. If you continue to modify the project after submitting it, the changes will not be available to users downloading your project through the store until this value is updated. If you don't know exactly how this works, don't worry about it. One of the Naomi core developers will add this value for you when your naomi-plugins pull request is merged.

Once you are finished adding your line, create a pull request to merge your changes back into the naomi-plugins repository. Once your pull request has been merged, users will be able to use Naomi's built-in store functions to install your plugin.