Rhasspy Natural Language Understanding

Library for parsing Rhasspy sentence templates, doing intent recognition, and generating ARPA language models.

Requirements

• Python 3.7

Installation

$ git clone https://github.com/rhasspy/rhasspy-nlu
$ cd rhasspy-nlu
$ ./configure
$ make
$ make install

Running

$ bin/rhasspy-nlu <ARGS>

Parsing Sentence Templates

Rhasspy voice commands are stored in text files formatted like this:

[Intent1]
this is a sentence
this is another sentence

[Intent2]
a sentence in a different intent

You can parse these into a structured representation with rhasspynlu.parse_ini and then convert them to a graph using rhasspynlu.intents_to_graph:

import rhasspynlu

# Load and parse
intents = rhasspynlu.parse_ini(
"""
[LightOn]
turn on [the] (living room lamp | kitchen light){name}
"""
)

graph = rhasspynlu.intents_to_graph(intents)

The result is a directed graph whose states are words and edges are input/output labels.

You can pass an intent_filter function to parse_ini to return True for only the intent names you want to parse. Additionally, a function can be provided for the sentence_transform argument that each sentence will be passed through (e.g., to lower case).

Template Syntax

Sentence templates are based on the JSGF standard. The following constructs are available:

• Optional words
  • this is [a] test - the word "a" may or may not be present
• Alternatives
  • set color to (red | green | blue) - either "red", "green", or "blue" is possible
• Tags
  • turn on the [den | playroom]{location} light - named entity location will be either "den" or "playroom"
• Substitutions
  • make ten:10 coffees - output will be "make 10 coffees"
  • turn off the: (television | tele):tv - output will be "turn off tv"
  • set brightness to (medium | half){brightness:50} - named entity brightness will be "50"
• Rules
  • rule_name = rule body can be referenced as <rule_name>
• Slots
  • $slot will be replaced by a list of sentences in the replacements argument of intents_to_graph

Rules

Named rules can be added to your template file using the syntax:

rule_name = rule body

and then reference using <rule_name>. The body of a rule is a regular sentence, which may itself contain references to other rules.

You can refrence rules from different intents by prefixing the rule name with the intent name and a dot:

[Intent1]
rule = a test
this is <rule>

[Intent2]
rule = this is
<rule> <Intent1.rule>

In the example above, Intent2 uses its local <rule> as well as the <rule> from Intent1.

Slots

Slot names are prefixed with a dollar sign ($). When calling intents_to_graph, the replacements argument is a dictionary whose keys are slot names (with $) and whose values are lists of (parsed) Sentence objects. Each $slot will be replaced by the corresponding list of sentences, which may contain optional words, tags, rules, and other slots.

For example:

import rhasspynlu

# Load and parse
intents = rhasspynlu.parse_ini(
"""
[SetColor]
set color to $color
"""
)

graph = rhasspynlu.intents_to_graph(
    intents, replacements = {
        "$color": [rhasspynlu.Sentence.parse("red | green | blue")]
    }
)

will replace $color with "red", "green", or "blue".

Intent Recognition

After converting your sentence templates to a graph, you can recognize sentences. Assuming you have a .ini file like this:

[LightOn]
turn on [the] (living room lamp | kitchen light){name}

You can recognize sentences with:

from pathlib import Path
import rhasspynlu

# Load and parse
intents = rhasspynlu.parse_ini(Path("sentences.ini"))
graph = rhasspynlu.intents_to_graph(intents)

rhasspynlu.recognize("turn on living room lamp", graph)

will return a list of Recognition objects like:

[
    Recognition(
        intent=Intent(name='LightOn', confidence=1.0),
        entities=[
            Entity(
                entity='name',
                value='living room lamp',
                raw_value='living room lamp',
                start=8,
                raw_start=8,
                end=24,
                raw_end=24,
                tokens=['living', 'room', 'lamp'],
                raw_tokens=['living', 'room', 'lamp']
            )
        ],
        text='turn on living room lamp',
        raw_text='turn on living room lamp',
        recognize_seconds=0.00010710899914556649,
        tokens=['turn', 'on', 'living', 'room', 'lamp'],
        raw_tokens=['turn', 'on', 'living', 'room', 'lamp']
    )
]

An empty list means that recognition has failed. You can easily convert Recognition objects to JSON:

...

import json

recognitions = rhasspynlu.recognize("turn on living room lamp", graph)
if recognitions:
    recognition_dict = recognitions[0].asdict()
    print(json.dumps(recognition_dict))

You can also pass an intent_filter function to recognize to return True only for intent names you want to include in the search.

Tokens

If your sentence is tokenized by something other than whitespace, pass the list of tokens into recognize instead of a string.

Recognition Fields

The rhasspynlu.Recognition object has the following fields:

• intent - a rhasspynlu.Intent instance
  • name - name of recognized intent
  • confidence - number for 0-1, 1 being sure
• text - substituted input text
• raw_text - input text
• entities - list of rhasspynlu.Entity objects
  • entity - name of recognized entity ("name" in (input:output){name})
  • value - substituted value of recognized entity ("output" in (input:output){name})
  • tokens - list of words in value
  • start - start index of value in text
  • end - end index of value in text (exclusive)
  • raw_value - value of recognized entity ("input" in (input:output){name})
  • raw_tokens - list of words in raw_value
  • raw_start - start index of raw_value in raw_text
  • raw_end - end index of raw_value in raw_text (exclusive)
• recognize_seconds - seconds taken for recognize

Stop Words

You can pass a set of stop_words to recognize:

rhasspynlu.recognize("turn on that living room lamp", graph, stop_words=set(["that"]))

Stop words in the input sentence will be skipped over if they don't match the graph.

Strict Recognition

For faster, but less flexible recognition, set fuzzy to False:

rhasspynlu.recognize("turn on the living room lamp", graph, fuzzy=False)

This is at least twice as fast, but will fail if the sentence is not precisely present in the graph.

Strict recognition also supports stop_words for a little added flexibility. If recognition without stop_words fails, a second attempt will be made using stop_words.

Converters

Value conversions can be applied during recognition, such as converting the string "10" to the integer 10. Following a word, sequence, or tag name with "!converter" will run "converter" on the string value during recognize:

import rhasspynlu

# Load and parse
intents = rhasspynlu.parse_ini(
"""
[SetBrightness]
set brightness to (one: hundred:100)!int
"""
)

graph = rhasspynlu.intents_to_graph(intents)

recognitions = rhasspynlu.recognize("set brightness to one hundred", graph)
assert recognitions[0].tokens[-1] == 100

Converters can be applied to tags/entities as well:

import rhasspynlu

# Load and parse
intents = rhasspynlu.parse_ini(
"""
[SetBrightness]
set brightness to (one:1 | two:2){value!int}
"""
)

graph = rhasspynlu.intents_to_graph(intents)

recognitions = rhasspynlu.recognize("set brightness to two", graph)
assert recognitions[0].tokens[-1] == 2

The following default converters are available in rhasspynlu:

• int - convert to integer
• float - convert to real
• bool - convert to boolean
• lower - lower-case
• upper - upper-case

You may override these converters by passing a dictionary to the converters argument of recognize. To supply additional converters (instead of overriding), use extra_converters:

import rhasspynlu

# Load and parse
intents = rhasspynlu.parse_ini(
"""
[SetBrightness]
set brightness to (one:1 | two:2){value!myconverter}
"""
)

graph = rhasspynlu.intents_to_graph(intents)

recognitions = rhasspynlu.recognize(
    "set brightness to two",
    graph,
    extra_converters={
        "myconverter": lambda *values: [int(v)**2 for v in values]
    }
)
assert recognitions[0].tokens[-1] == 4

Lastly, you can chain converters together with multiple "!":

import rhasspynlu

# Load and parse
intents = rhasspynlu.parse_ini(
"""
[SetBrightness]
set brightness to (one:1 | two:2){value!int!cube}
"""
)

graph = rhasspynlu.intents_to_graph(intents)

recognitions = rhasspynlu.recognize(
    "set brightness to two",
    graph,
    extra_converters={
        "cube": lambda *values: [v**3 for v in values]
    }
)
assert recognitions[0].tokens[-1] == 8

ARPA Language Models

You can compute ngram counts from a rhasspynlu graph, useful for generating ARPA language models. These models can be used by speech recognition systems, such as Pocketsphinx, Kaldi, and Julius.

import rhasspynlu

# Load and parse
intents = rhasspynlu.parse_ini(
"""
[SetColor]
set light to (red | green | blue)
"""
)

graph = rhasspynlu.intents_to_graph(intents)
counts = rhasspynlu.get_intent_ngram_counts(
    graph,
    pad_start="<s>",
    pad_end="</s>",
    order=3
)

# Print counts by intent
for intent_name in counts:
    print(intent_name)
    for ngram, count in counts[intent_name].items():
        print(ngram, count)

    print("")

will print something like:

SetColor
('<s>',) 3
('set',) 3
('<s>', 'set') 3
('light',) 3
('set', 'light') 3
('<s>', 'set', 'light') 3
('to',) 3
('light', 'to') 3
('set', 'light', 'to') 3
('red',) 1
('to', 'red') 1
('light', 'to', 'red') 1
('green',) 1
('to', 'green') 1
('light', 'to', 'green') 1
('blue',) 1
('to', 'blue') 1
('light', 'to', 'blue') 1
('</s>',) 3
('red', '</s>') 1
('green', '</s>') 1
('blue', '</s>') 1
('to', 'red', '</s>') 1
('to', 'green', '</s>') 1
('to', 'blue', '</s>') 1

Opengrm

If you have the Opengrm command-line tools in your PATH, you can use rhasspynlu to generate language models in the ARPA format.

The graph_to_fst and fst_to_arpa functions are used to convert between formats. Calling fst_to_arpa requires the following binaries to be present in your PATH:

• fstcompile (from OpenFST)
• ngramcount
• ngrammake
• ngrammerge
• ngramprint
• ngramread

Example:

# Convert to FST
graph_fst = rhasspynlu.graph_to_fst(graph)

# Write FST and symbol text files
graph_fst.write("my_fst.txt", "input_symbols.txt", "output_symbols.txt")

# Compile and convert to ARPA language model
rhasspynlu.fst_to_arpa(
    "my_fst.txt",
    "input_symbols.txt",
    "output_symbols.txt",
    "my_arpa.lm"
)

You can now use my_arpa.lm in any speech recognizer that accepts ARPA-formatted language models.

Language Model Mixing

If you have an existing language model that you'd like to mix with Rhasspy voice commands, you will first need to convert it to an FST:

rhasspynlu.fst_to_arpa("existing_arpa.lm", "existing_arpa.fst")

Now when you call fst_to_arpa, make sure to provide the base_fst_weight argument. This is a tuple with the path to your existing ARPA FST and a mixture weight between 0 and 1. A weight of 0.05 means that the base language model will receive 5% of the overall probability mass in the language model. The rest of the mass will be given to your custom voice commands.

Example:

rhasspynlu.fst_to_arpa(
    "my_fst.txt",
    "input_symbols.txt",
    "output_symbols.txt",
    "my_arpa.lm",
    base_fst_weight=("existing_arpa.fst", 0.05)
)

Command Line Usage

The rhasspynlu module can be run directly to convert sentences.ini files into JSON graphs or FST text files:

python3 -m rhasspynlu sentences.ini > graph.json

You can pass multiple .ini files as arguments, and they will be combined. Adding a --fst argument will write out FST text files instead:

python3 -m rhasspynlu sentences.ini --fst

This will output three files in the current directory:

• fst.txt - finite state transducer as text
• fst.isymbols.txt - input symbols
• fst.osymbols.txt - output symbols

These file names can be changed with the --fst-text, --fst-isymbols, and --fst-osymbols arguments, respectively.

Compile to a binary FST using fstcompile (from OpenFST) with:

fstcompile \
    --isymbols=fst.isymbols.txt \
    --osymbols=fst.osymbols.txt \
    --keep_isymbols=1 \
    --keep_osymbols=1 \
    fst.txt \
    out.fst

Word Pronunciations

Methods for loading and using phonetic pronunciation dictionaries are provided in rhasspynlu.g2p ("g2p" stands for "grapheme to phoneme").

Dictionaries are expected in the same format as the CMU Pronouncing Dictionary, which is simply one word per line with whitespace separating words and phonemes:

yawn Y AO N
test T EH S T
say S EY
who HH UW
bee B IY
azure AE ZH ER
read R EH D
read(2) R IY D

When multiple pronunciations are available for a word (like "read" in the previous example), a (N) can be suffixed to the word.

You can load a phonetic dictionary into a Python dictionary with rhasspynlu.g2p.read_pronunciations:

import rhasspynlu.g2p

with open("/path/to/file.dict", "r") as dict_file:
    pronunciations = rhasspynlu.g2p.read_pronunciations(dict_file)

assert pronunciations == {
    "yawn": [["Y", "AO", "N"]],
    "test": [["T", "EH", "S", "T"]],
    "say": [["S", "EY"]],
    "who": [["HH", "UW"]],
    "bee": [["B", "IY"]],
    "azure": [["AE", "ZH", "ER"]],
    "read": [["R", "EH", "D"], ["R", "IY", "D"]],
}

See voice2json profiles for pre-built phonetic dictionaries.

Guessing Pronunciations

The rhasspynlu.g2p.guess_pronunciations function uses Phonetisaurus and a pre-trained grapheme to phoneme model to guess pronunciations for unknown words. You will need phonetisaurus-apply in your $PATH and the pre-trained model (g2p.fst) available:

import rhasspynlu.g2p

guesses = rhasspynlu.g2p.guess_pronunciations(
    ["moogle", "ploop"], "/path/to/g2p.fst", num_guesses=1
)

print(list(guesses))

# Something like: [
#   ('moogle', ['M', 'UW', 'G', 'AH', 'L']),
#   ('ploop', ['P', 'L', 'UW', 'P'])
# ]

See voice2json profiles for pre-trained g2p models.

Sounds Like Pronunciations

Rhasspy NLU supports an alternative way of specifying word pronunciations. Instead of specifying phonemes directly, you can describe how a word should be pronounced by referencing other words:

unknown_word1 known_word1 [known_word2] ...
...

For example, the singer Beyoncé sounds like a combination of the words "bee yawn say":

beyoncé bee yawn say

The rhasspynlu.g2p.load_sounds_like function will parse this text and, when given an existing pronunciation dictionary, generate a new pronunciation:

import io

import rhasspynlu.g2p

# Load existing dictionary
pronunciations = rhasspynlu.g2p.read_pronunciations("/path/to/file.dict")

sounds_like = """
beyoncé bee yawn say
"""

with io.StringIO(sounds_like) as f:
    rhasspynlu.g2p.load_sounds_like(f, pronunciations)

print(pronunciations["beyoncé"])

# Something like: [['B', 'IY', 'Y', 'AO', 'N', 'S', 'EY']]

You may reference a specific pronunciation for a known word using the word(N) syntax, where N is 1-based. Pronunciations are loaded in line order, so the order is predictable. For example, read(2) will reference the second pronunciation of the word "read". Without an (N), all pronunciations found will be used.

Phoneme Literals

You can interject phonetic chunks into these pronunciations too. For example, the word "hooiser" sounds like "who" and the "-zure" in "azure":

hooiser who /Z 3/

Text between slashes (/) will be interpreted as phonemes in the configured speech system.

Word Segments

If a grapheme-to-phoneme alignment corupus is available, segments of words can also be used for pronunciations. Using the "hooiser" example above, we can replace the phonemes with:

hooiser who a>zure<

This will combine the pronunciation of "who" from the current phonetic dictionaries (base_dictionary.txt and custom_words.txt) and the "-zure" from the word "azure".

The brackets point >at< the segment of the word that you want to contribute to the pronunciation. This is accomplished using a grapheme-to-phoneme alignment corpus generated with phonetisaurus and a pre-built phonetic dictionary. In the a>zure< example, the word "azure" is located in the alignment corpus, and the output phonemes from the phonemes "zure" in it are used.

import io

import rhasspynlu.g2p

# Load existing dictionary
pronunciations = rhasspynlu.g2p.read_pronunciations("/path/to/file.dict")

# Example alignment corpus:
# a}AE z}ZH u|r}ER e}_
alignment = rhasspynlu.g2p.load_g2p_corpus("/path/to/g2p.corpus")

sounds_like = """
hooiser who a>zure<
"""

with io.StringIO(sounds_like) as f:
    rhasspynlu.g2p.load_sounds_like(
        f, pronunciations, g2p_alignment=alignment
    )

print(pronunciations["hooiser"])

# Something like [["HH", "UW", "ZH", "ER"]]

See voice2json profiles for g2p alignment corpora.