/snowboy

DNN based hotword and wake word detection toolkit

Primary LanguageMakefile

Snowboy Hotword Detection

by KITT.AI.

Home Page

Full Documentation and FAQ

Discussion Group (or send email to snowboy-discussion@kitt.ai)

(The discussion group is new since September 2016 as we are getting many messages every day. Please send general questions there. For bugs, use Github issues.)

Version: 1.2.0 (3/25/2017)

Alexa support

Snowboy now brings hands-free experience to the Alexa AVS sample app on Raspberry Pi! See more info below regarding the performance and how you can use other hotword models.

Performance

The performance of hotword detection usually depends on the actually environment, e.g., is it used with a quality microphone, is it used on the street, in a kitchen, or is there any background noise, etc. So we feel it is best for the users to evaluate it in their real environment. For the evaluation purpose, we have prepared an Android app which can be installed and run out of box: SnowboyAlexaDemo.apk (please uninstall any previous one first if you installed this app before).

Personal model

# Please replace YOUR_PERSONAL_MODEL.pmdl with the personal model you just
# created, and $ALEXA_AVS_SAMPLE_APP_PATH with the actual path where you
# cloned the Alexa AVS sample app repository.
cp YOUR_PERSONAL_MODEL.pmdl $ALEXA_AVS_SAMPLE_APP_PATH/samples/wakeWordAgent/ext/resources/alexa.umdl
# Please replace $ALEXA_AVS_SAMPLE_APP_PATH with the actual path where you
# cloned the Alexa AVS sample app repository.
cd $ALEXA_AVS_SAMPLE_APP_PATH/samples/wakeWordAgent/src/

# Modify KittAiSnowboyWakeWordEngine.cpp and update SENSITIVITY at line 28.
# Modify KittAiSnowboyWakeWordEngine.cpp and set APPLY_FRONTEND to false at
# line 30.
make
  • Run the wake word agent with engine set to kitt_ai!

Universal model

# Please replace YOUR_UNIVERSAL_MODEL.umdl with the personal model you just
# created, and $ALEXA_AVS_SAMPLE_APP_PATH with the actual path where you
# cloned the Alexa AVS sample app repository.
cp YOUR_UNIVERSAL_MODEL.umdl $ALEXA_AVS_SAMPLE_APP_PATH/samples/wakeWordAgent/ext/resources/alexa.umdl
# Please replace $ALEXA_AVS_SAMPLE_APP_PATH with the actual path where you
# cloned the Alexa AVS sample app repository.
cd $ALEXA_AVS_SAMPLE_APP_PATH/samples/wakeWordAgent/src/

# Modify KittAiSnowboyWakeWordEngine.cpp and update SENSITIVITY at line 28.
make
  • Run the wake word agent with engine set to kitt_ai!

Hotword as a Service

Snowboy now offers Hotword as a Service through the https://snowboy.kitt.ai/api/v1/train/ endpoint. Check out the Full Documentation and example Python/Bash script (other language contributions are very welcome).

As a quick start, POST to https://snowboy.kitt.ai/api/v1/train:

{
    "name": "a word",
    "language": "en",
    "age_group": "10_19",
    "gender": "F",
    "microphone": "mic type",
    "token": "<your auth token>",
    "voice_samples": [
        {wave: "<base64 encoded wave data>"},
        {wave: "<base64 encoded wave data>"},
        {wave: "<base64 encoded wave data>"}
    ]
}

then you'll get a trained personal model in return!

Introduction

Snowboy is a customizable hotword detection engine for you to create your own hotword like "OK Google" or "Alexa". It is powered by deep neural networks and has the following properties:

  • highly customizable: you can freely define your own magic phrase here – let it be “open sesame”, “garage door open”, or “hello dreamhouse”, you name it.

  • always listening but protects your privacy: Snowboy does not use Internet and does not stream your voice to the cloud.

  • light-weight and embedded: it even runs on a Raspberry Pi and consumes less than 10% CPU on the weakest Pi (single-core 700MHz ARMv6).

  • Apache licensed!

Currently Snowboy supports (look into the lib folder):

  • all versions of Raspberry Pi (with Raspbian based on Debian Jessie 8.0)
  • 64bit Mac OS X
  • 64bit Ubuntu (12.04 and 14.04)
  • iOS
  • Android
  • Pine64 (Debian Jessie 8.5, 3.10.102 BSP2)
  • Nvidia Jetson TX1 (use above Pine64 package)
  • Nvidia Jetson TX2 (use above Pine64 package)
  • Intel Edison (Ubilinux based on Debian Wheezy 7.8)
  • Samsung Artik (built with Fedora 25 for ARMv7)
  • ARM64 (aarch64, Ubuntu 16.04)

It ships in the form of a C++ library with language-dependent wrappers generated by SWIG. We welcome wrappers for new languages -- feel free to send a pull request!

Currently we have built wrappers for:

  • Java/Android
  • Go (thanks to @brentnd)
  • Node (thanks to @evancohen and @nekuz0r)
  • Perl (thanks to @iboguslavsky)
  • Python
  • iOS/Swift3 (thanks to @grimlockrocks)
  • iOS/Object-C (thanks to @patrickjquinn)

If you want support on other hardware/OS, please send your request to snowboy@kitt.ai

Note: Snowboy does not support Windows yet. Please build Snowboy on *nix platforms.

Pricing for Snowboy models

Hackers: free

  • Personal use
  • Community support

Business: please contact us at snowboy@kitt.ai

  • Personal use
  • Commercial license
  • Technical support

Pretrained universal models

We provide pretrained universal models for testing purpose. When you test those models, bear in mind that they may not be optimized for your specific device or environment.

Here is the list of the models, and the parameters that you have to use for them:

  • resources/snowboy.umdl: Universal model for the hotword "Snowboy". Set SetSensitivity to 0.5 for better performance.
  • resources/alexa.umdl: Universal model for the hotword "Alexa". Set SetSensitivity to 0.5, and preferably set ApplyFrontend (only works on Raspberry Pi) to true. This model is depressed.
  • resources/alexa/alexa_02092017.umdl: Universal model for the hotword "Alexa". This is still work in progress. Set SetSensitivity to 0.15.
  • resources/alexa/alexa-avs-sample-app/alexa.umdl: Universal model for the hotword "Alexa" optimized for Alexa AVS sample app. Set SetSensitivity to 0.6, and set ApplyFrontend (only works on Raspberry Pi) to true. This is so far the best "Alexa" model we released publicly, when ApplyFrontend is set to true.

Precompiled node module

Snowboy is available in the form of a native node module precompiled for: 64 bit Ubuntu, MacOS X, and the Raspberry Pi (Raspbian 8.0+). For quick installation run:

npm install --save snowboy

For sample usage see the examples/Node folder. You may have to install dependencies like fs, wav or node-record-lpcm16 depending on which script you use.

Precompiled Binaries with Python Demo

  • 64 bit Ubuntu 12.04 / 14.04
  • MacOS X
  • Raspberry Pi with Raspbian 8.0, all versions (1/2/3/Zero)
  • Pine64 (Debian Jessie 8.5 (3.10.102)), Nvidia Jetson TX1 and Nvidia Jetson TX2 (download)
  • Intel Edison (Ubilinux based on Debian Wheezy 7.8) (download)

If you want to compile a version against your own environment/language, read on.

Dependencies

To run the demo you will likely need the following, depending on which demo you use and what platform you are working with:

  • SoX (audio conversion)
  • PortAudio or PyAudio (audio capturing)
  • SWIG 3.0.10 or above (compiling Snowboy for different languages/platforms)
  • ATLAS or OpenBLAS (matrix computation)

You can also find the exact commands you need to install the dependencies on Mac OS X, Ubuntu or Raspberry Pi below.

Mac OS X

brew install swig, sox, portaudio and its Python binding pyaudio:

brew install swig portaudio sox
pip install pyaudio

If you don't have Homebrew installed, please download it here. If you don't have pip, you can install it here.

Make sure that you can record audio with your microphone:

rec t.wav

Ubuntu/Raspberry Pi/Pine64/Nvidia Jetson TX1/Nvidia Jetson TX2

First apt-get install swig, sox, portaudio and its Python binding pyaudio:

sudo apt-get install swig3.0 python-pyaudio python3-pyaudio sox
pip install pyaudio

Then install the atlas matrix computing library:

sudo apt-get install libatlas-base-dev

Make sure that you can record audio with your microphone:

rec t.wav

If you need extra setup on your audio (especially on a Raspberry Pi), please see the full documentation.

Compile a Node addon

Compiling a node addon for Linux and the Raspberry Pi requires the installation of the following dependencies:

sudo apt-get install libmagic-dev libatlas-base-dev

Then to compile the addon run the following from the root of the snowboy repository:

npm install
node-pre-gyp clean configure build

Compile a Java Wrapper

# Make sure you have JDK installed.
cd swig/Java
make

SWIG will generate a directory called java which contains converted Java wrappers and a directory called jniLibs which contains the JNI library.

To run the Java example script:

cd examples/Java
make run

Compile a Python Wrapper

cd swig/Python
make

SWIG will generate a _snowboydetect.so file and a simple (but hard-to-read) python wrapper snowboydetect.py. We have provided a higher level python wrapper snowboydecoder.py on top of that.

Feel free to adapt the Makefile in swig/Python to your own system's setting if you cannot make it.

Compile a GO Wrapper

cd examples/Go
go get github.com/Kitt-AI/snowboy/swig/Go
go build -o snowboy main.go
./snowboy ../../resources/snowboy.umdl ../../resources/snowboy.wav

Expected Output:

Snowboy detecting keyword in ../../resources/snowboy.wav
Snowboy detected keyword  1

For more, please read examples/Go/readme.md.

Compile a Perl Wrapper

cd swig/Perl
make

The Perl examples include training personal hotword using the KITT.AI RESTful APIs, adding Google Speech API after the hotword detection, etc. To run the examples, do the following

cd examples/Perl

# Install cpanm, if you don't already have it.
curl -L https://cpanmin.us | perl - --sudo App::cpanminus

# Install the dependencies. Note, on Linux you will have to install the
# PortAudio package first, using e.g.:
# apt-get install portaudio19-dev
sudo cpanm --installdeps .

# Run the unit test.
./snowboy_unit_test.pl

# Run the personal model training example.
./snowboy_RESTful_train.pl <API_TOKEN> <Hotword> <Language>

# Run the Snowboy Google Speech API example. By default it uses the Snowboy
# universal hotword.
./snowboy_googlevoice.pl <Google_API_Key> [Hotword_Model]

Compile an iOS Wrapper

Using Snowboy library in Objective-C does not really require a wrapper. It is basically the same as using C++ library in Objective-C. We have compiled a "fat" static library for iOS devices, see the library here lib/ios/libsnowboy-detect.a.

To initialize Snowboy detector in Objective-C:

snowboy::SnowboyDetect* snowboyDetector = new snowboy::SnowboyDetect(
    std::string([[[NSBundle mainBundle]pathForResource:@"common" ofType:@"res"] UTF8String]),
    std::string([[[NSBundle mainBundle]pathForResource:@"snowboy" ofType:@"umdl"] UTF8String]));
snowboyDetector->SetSensitivity("0.45");        // Sensitivity for each hotword
snowboyDetector->SetAudioGain(2.0);             // Audio gain for detection

To run hotword detection in Objective-C:

int result = snowboyDetector->RunDetection(buffer[0], bufferSize);  // buffer[0] is a float array

You may want to play with the frequency of the calls to RunDetection(), which controls the CPU usage and the detection latency.

Thanks to @patrickjquinn and @grimlockrocks, we now have examples of using Snowboy in both Objective-C and Swift3. Check out the examples at examples/iOS/, and the screenshots below!

Obj-C Example Swift3 Example

Compile an Android Wrapper

Full README and tutorial is in Android README and here's a screenshot:

Android Alexa Demo

We have prepared an Android app which can be installed and run out of box: SnowboyAlexaDemo.apk (please uninstall any previous one first if you installed this app before).

Quick Start for Python Demo

Go to the examples/Python folder and open your python console:

In [1]: import snowboydecoder

In [2]: def detected_callback():
   ....:     print "hotword detected"
   ....:

In [3]: detector = snowboydecoder.HotwordDetector("resources/snowboy.umdl", sensitivity=0.5, audio_gain=1)

In [4]: detector.start(detected_callback)

Then speak "snowboy" to your microphone to see whetheer Snowboy detects you.

The snowboy.umdl file is a "universal" model that detect different people speaking "snowboy". If you want other hotwords, please go to snowboy.kitt.ai to record, train and downloand your own personal model (a .pmdl file).

When sensitiviy is higher, the hotword gets more easily triggered. But you might get more false alarms.

audio_gain controls whether to increase (>1) or decrease (<1) input volume.

Two demo files demo.py and demo2.py are provided to show more usages.

Note: if you see the following error:

TypeError: __init__() got an unexpected keyword argument 'model_str'

You are probably using an old version of SWIG. Please upgrade. We have tested with SWIG version 3.0.7 and 3.0.8.

Advanced Usages & Demos

See Full Documentation.

Change Log

v1.2.0, 3/25/2017

  • Added better Alexa model for Alexa AVS sample app
  • New decoder that works well for short hotwords like Alexa

v1.1.1, 3/24/2017

  • Added Android demo
  • Added iOS demos
  • Added Samsung Artik support
  • Added Go support
  • Added Intel Edison support
  • Added Pine64 support
  • Added Perl Support
  • Added a more robust "Alexa" model (umdl)
  • Offering Hotword as a Service through /api/v1/train endpoint.
  • Decoder is not changed.

v1.1.0, 9/20/2016

  • Added library for Node.
  • Added support for Python3.
  • Added universal model alexa.umdl
  • Updated universal model snowboy.umdl so that it works in noisy environment.

v1.0.4, 7/13/2016

  • Updated universal snowboy.umdl model to make it more robust.
  • Various improvements to speed up the detection.
  • Bug fixes.

v1.0.3, 6/4/2016

  • Updated universal snowboy.umdl model to make it more robust in non-speech environment.
  • Fixed bug when using float as input data.
  • Added library support for Android ARMV7 architecture.
  • Added library for iOS.

v1.0.2, 5/24/2016

  • Updated universal snowboy.umdl model
  • added C++ examples, docs will come in next release.

v1.0.1, 5/16/2016

  • VAD now returns -2 on silence, -1 on error, 0 on voice and >0 on triggered models
  • added static library for Raspberry Pi in case people want to compile themselves instead of using the binary version

v1.0.0, 5/10/2016

  • initial release