Voice Command Lifecycle

Python library to manage the life-cycle of voice commands. Useful working with Alexa Voice Service.

---

Installation

pip install command_lifecycle

Wakeword detector

A wakeword is a specific word that triggers the code to spring into action. It allows your code to be idle until the specific word is uttered.

The audio lifecycle uses snowboy to determine if the wakeword was uttered. The library will need to be installed first.

Once you have compiled snowboy, copy the compiled snowboy folder to the top level of you project. By default, the folder structure should be:

.
├── ...
├── snowboy
|   ├── snowboy-detect-swig.cc
|   ├── snowboydetect.py
|   └── resources
|       ├── alexa.umdl
|       └── common.res
└── ...

If the default structure does not suit your needs can customize the wakeword detector.

Usage

You should send a steady stream of audio to to the lifecycle by repetitively calling lifecycle.extend_audio(some_audio_bytes). If the wakeword such as "Alexa" (default), or "ok, Google" was uttered then handle_command_started is called. handle_command_finised is then called once the command audio that followed the wakeword has finished.

Microphone audio

import pyaudio

import command_lifecycle


class AudioLifecycle(command_lifecycle.BaseAudioLifecycle):

    def handle_command_started(self, wakeword_name):
        super().handle_command_started(wakeword_name)
        print(f'The audio contained {wakeword_name}!')

    def handle_command_finised(self):
        super().handle_command_finised()
        print('The command has finished')

lifecycle = AudioLifecycle()

p = pyaudio.PyAudio()
stream = p.open(format=pyaudio.paInt16, channels=1, rate=16000, input=True)

try:
    print('listening. Start by saying "Alexa". Press CTRL + C to exit.')
    while True:
        lifecycle.extend_audio(stream.read(1024))
finally:
    stream.stop_stream()
    stream.close()
    p.terminate()

File audio

import wave

import command_lifecycle

class AudioLifecycle(command_lifecycle.BaseAudioLifecycle):
    def handle_command_started(self, wakeword_name):
        super().handle_command_started(wakeword_name)
        print(f'The audio contained {wakeword_name}!')


lifecycle = AudioLifecycle()
with wave.open('./tests/resources/alexa_what_time_is_it.wav', 'rb') as f:
    while f.tell() < f.getnframes():
        lifecycle.extend_audio(f.readframes(1024))

print('The command has finished')

Usage with Alexa

command_lifecycle is useful for interacting with voice services. The lifecycle waits until a wakeword was issued and then start streaming the audio command to the voice service (using Alexa Voice Service Client), then do something useful with the response:

from avs_client.avs_client.client import AlexaVoiceServiceClient
import pyaudio

import command_lifecycle


class AudioLifecycle(command_lifecycle.BaseAudioLifecycle):
    alexa_client = AlexaVoiceServiceClient(
        client_id='my-client-id'
        secret='my-secret',
        refresh_token='my-refresh-token',
    )

    def __init__(self):
        self.alexa_client.connect()
        super().__init__()

    def handle_command_started(self, wakeword_name):
        super().handle_command_started(wakeword_name)
        audio_file = command_lifecycle.to_audio_file()
        for directive in self.alexa_client.send_audio_file(audio_file):
            # do something with the AVS audio response, e.g., play it.

lifecycle = AudioLifecycle()

p = pyaudio.PyAudio()
stream = p.open(format=pyaudio.paInt16, channels=1, rate=16000, input=True)

try:
    print('listening. Start by saying "Alexa". Press CTRL + C to exit.')
    while True:
        lifecycle.extend_audio(stream.read(1024))
finally:
    stream.stop_stream()
    stream.close()
    p.terminate()

Customization

Wakeword

The default wakeword is "Alexa". This can be changed by sub-classing command_lifecycle.wakeword.SnowboyWakewordDetector:

from command_lifecycle import wakeword


class MySnowboyWakewordDetector(wakeword.SnowboyWakewordDetector):
    decoder_models = [
        {
            'name': 'CUSTOM',
            'model': b'path/to/custom-wakeword-model.umdl'
            'sensitivity': b'0.5',
        }
    ]


class AudioLifecycle(lifecycle.BaseAudioLifecycle):
    audio_detector_class = MySnowboyWakewordDetector

    def handle_command_started(self, wakeword_name):
        super().handle_command_started(wakeword_name)
        print(f'The audio contained the {wakeword_name}!')

    def handle_command_finised(self):
        super().handle_command_finised()
        print('The command has finished')

lifecycle = AudioLifecycle()

# now load the audio into lifecycle

See the Snowboy docs for steps on creating custom wakeword models.

Multiple Wakewords

Triggering different behaviour for different wakeword may be desirable. To do this use multiple items in decoder_models:

from command_lifecycle import wakeword


class MyMultipleWakewordDetector(wakeword.SnowboyWakewordDetector):
    GOOGLE = 'GOOGLE'

    decoder_models = wakeword.SnowboyWakewordDetector.decoder_models + [
        {
            'name': GOOGLE,
            'model': b'path/to/okgoogle.umdl',
            'sensitivity': b'0.5',
        }
    ]


class AudioLifecycle(lifecycle.BaseAudioLifecycle):
    audio_detector_class = MyMultipleWakewordDetector

    def handle_command_started(self, wakeword_name):
        if wakeword_name == self.audio_detector.ALEXA:
            print('Alexa standing by')
        elif wakeword_name == self.audio_detector.GOOGLE:
            print('Google at your service')
        super().handle_command_started(wakeword_name)

You can download wakewords from here.

Wakeword detector

Snowboy is the default wakeword detector. Other wakeword detectors can be used by sub-classing command_lifecycle.wakeword.BaseWakewordDetector and setting wakeword_detector_class to your custom class:

import wave

from command_lifecycle import lifecycle, wakeword


class MyCustomWakewordDetector(wakeword.BaseWakewordDetector):
    import_error_message = 'Cannot import wakeword library!'
    wakeword_library_import_path = 'path.to.wakeword.Library'

    def was_wakeword_uttered(self, buffer):
        # use the library to check if the audio in the buffer has the wakeword.
        # not `buffer.get()` returns the audio inside the buffer.
        ...

    def is_talking(self, buffer):
        # use the library to check if the audio in the buffer has audible words
        # not `buffer.get()` returns the audio inside the buffer.
        ...


class AudioLifecycle(lifecycle.BaseAudioLifecycle):
    audio_detector_class = MyCustomWakewordDetector
    ...


lifecycle = AudioLifecycle()
# now load the audio into lifecycle

Handling input data

Three input data formats are supported:

Converter	Notes
NoOperationConverter	default Input data is already wav bytes.
WavIntSamplestoWavConverter	Input data is list of integers.
WebAudioToWavConverter	Input data is list of floats generated by a web browser.

Customize this by setting the lifecycle's audio_converter_class:

from command_lifecycle.helpers import WebAudioToWavConverter

class AudioLifecycle(lifecycle.BaseAudioLifecycle):
    audio_converter_class = WebAudioToWavConverter

Expecting slower or faster commands

The person giving the audio command might take a moment to collect their thoughts before finishing the command. This silence could be interpreted as the command ending, resulting in handle_command_finised being called prematurely.

To avoid this the lifecycle tolerates some silence in the command before the lifecycle timesout the command. This silence can happen at the beginning or middle of the command. Note a side-effect of this is there will be a pause between when the person has stopped talking and when handle_command_finised is called.

To change this default behaviour timeout_manager_class can be changed. The available timeout managers are:

Timeout manager	Notes
ShortTimeoutManager	Allows one second of silence.
MediumTimeoutManager	default Allows 2 seconds of silence.
LongTimeoutManager	Allows three seconds of silence.

To make a custom timeout manager create a subclass of command_lifecycle.timeout.BaseTimeoutManager:

import wave

from command_lifecycle import timeout, wakeword


class MyCustomTimeoutManager(timeout.BaseTimeoutManager):
    allowed_silent_seconds = 4


class AudioLifecycle(lifecycle.BaseAudioLifecycle):
    timeout_manager_class = MyCustomTimeoutManager

Unit test

To run the unit tests, call the following commands:

make test_requirements
make test

Versioning

We use SemVer for versioning. For the versions available, see the PyPI.

Other projects

This library is used by alexa-browser-client, which allows you to talk to Alexa from your browser.