Picovoice.ai

There is a great video for getting started:

I’ve distilled the commands out for you…

First pull out the repository – this took a long time for me… its BIG (4.2G at the time of this writing):

git clone https://github.com/Picovoice/Porcupine.git

Now that you have the repo.  There are a good number of hotword models to choose from.  Look in:

~/Porcupine/resources/keyword_files/raspberrypi

On my system:

americano_raspberrypi.ppn       caterpillar_raspberrypi.ppn       francesca_raspberrypi.ppn         picovoice_raspberrypi.ppn       terminator_raspberrypi.ppn
americano_raspberrypi_tiny.ppn  caterpillar_raspberrypi_tiny.ppn  francesca_raspberrypi_tiny.ppn    picovoice_raspberrypi_tiny.ppn  terminator_raspberrypi_tiny.ppn
avocado_raspberrypi.ppn         christina_raspberrypi.ppn         grapefruit_raspberrypi.ppn        pineapple_raspberrypi.ppn       vancouver_raspberrypi.ppn
avocado_raspberrypi_tiny.ppn    christina_raspberrypi_tiny.ppn    grapefruit_raspberrypi_tiny.ppn   pineapple_raspberrypi_tiny.ppn  vancouver_raspberrypi_tiny.ppn
blueberry_raspberrypi.ppn       dragonfly_raspberrypi.ppn         grasshopper_raspberrypi.ppn       porcupine_raspberrypi.ppn
blueberry_raspberrypi_tiny.ppn  dragonfly_raspberrypi_tiny.ppn    grasshopper_raspberrypi_tiny.ppn  porcupine_raspberrypi_tiny.ppn
bumblebee_raspberrypi.ppn       flamingo_raspberrypi.ppn          iguana_raspberrypi.ppn            raspberry_raspberrypi.ppn
bumblebee_raspberrypi_tiny.ppn  flamingo_raspberrypi_tiny.ppn     iguana_raspberrypi_tiny.ppn       raspberry_raspberrypi_tiny.ppn

Which means I can use these words as my “hotword”:

americano
avocado
blueberry
bumblebee
caterpillar
christina
dragonfly
flamingo
francescagrapefruit
grasshopper
iguana
picovoice
pineapple
porcupine
raspberry
terminator
vancouver

Let’s try “avocado“.  That’s kinda fun.  Now… next thing you’ll need is someway to get audio into the pi.  I like these dongles from Adafruit:

Ok, plug those in, and make sure they are working. A tool I like to use for recording a test audio sample, run this, then say “terminator” a couple of times:

arecord -D plughw:CARD=Device,DEV=0 test.wav

That should create an audio file called “test.wav” that you can play back to hear your sound quality:

aplay -D plughw:CARD=Device,DEV=0 test.wav

Picovoice doesn’t need awesome sound quality, its pretty danged robust. So… lets give it a whirl… but hold-on you need some python pre-reqs:

sudo apt-get update
sudo apt-get upgrade
sudo apt-get install build-essential libssl-dev libffi-dev python-dev
sudo apt-get install python-pip
sudo pip install enum34
# These two steps take a long time on the little pi
sudo pip install numpy
pip install -U numpy 
sudo pip install sndfile
sudo apt-get install python-pyaudio python3-pyaudio
sudo pip install soundfile
mkdir soundstuffs
cd soundstuffs
wget http://www.mega-nerd.com/libsndfile/files/libsndfile-1.0.28.tar.gz
tar xvf libsndfile-1.0.28.tar.gz
cd libsndfile-1.0.28
./configure
make
sudo make install
ldconfig

Phew! Keep going…

Training your own model… no not now… later…

To train your own model – you need to use a bigger computer.  Something like a mac or a PC.  Sadly, they don’t provide the tools on the PI.  

To tell porcupine to train on the word of your choosing.  In this example “computer” is the word:

cd Procupine
tools/optimizer/linux/x86_64/pv_porcupine_optimizer -r resources/ -p linux -o . -w "computer"

Let’s try it out!!!

The instructions I’m cribbing from are found here.

First, change directory into the Porcupine repo:

cd Porcupine

Then run the wake word command:

python demo/python/porcupine_demo.py --keyword_file_paths resources/keyword_files/raspberrypi/avocado_raspberrypi.ppn --library_path ./lib/raspberry-pi/cortex-a53/libpv_porcupine.so

Ok, it probably blew up with something like “Invalid Sample Rate”:

IOError: [Errno -9997] Invalid sample rate

That’s because the sample rate of your hardware is locked to something this software can’t handle.  You need something more flexible between you and the microphone.  

Alsa to the rescue!

Alsa has a default software layer sink that allows you to sample at any rate you need, and it just magically works.  To get at that magic, you have to create this new device.  So, create an “.asoundrc” file in your home directory, and put this in it:

pcm.!default {     
  capture.pcm "usb_mic"
}

pcm.usb_mic {
  type plug
  slave {
    pcm "hw:1,0"
  }
}

That configuration instructs Alsa to create the new device and call it “usb_mic”…. and make it default.

Now, run the command from before but with a command that lets you see the mics:

python ./demo/python/porcupine_demo.py --show_audio_devices_info

Take a look at the list… you should now see your new device:

'index': '0', 'name': 'bcm2835 ALSA: - (hw:0,0)', 'defaultSampleRate': '44100.0', 'maxInputChannels': '0'
'index': '1', 'name': 'bcm2835 ALSA: IEC958/HDMI (hw:0,1)', 'defaultSampleRate': '44100.0', 'maxInputChannels': '0'
'index': '2', 'name': 'USB Audio Device: - (hw:1,0)', 'defaultSampleRate': '44100.0', 'maxInputChannels': '1'
'index': '3', 'name': 'sysdefault', 'defaultSampleRate': '44100.0', 'maxInputChannels': '0'
'index': '4', 'name': 'usb_mic', 'defaultSampleRate': '44100.0', 'maxInputChannels': '128'
'index': '5', 'name': 'dmix', 'defaultSampleRate': '48000.0', 'maxInputChannels': '0'
'index': '6', 'name': 'default', 'defaultSampleRate': '44100.0', 'maxInputChannels': '32'

That “usb_mic” is the one you just created!  Now instruct Porcupine to use it with the command line parameter:

python demo/python/porcupine_demo.py --keyword_file_paths resources/keyword_files/raspberrypi/avocado_raspberrypi.ppn --library_path ./lib/raspberry-pi/cortex-a53/libpv_porcupine.so --input_audio_device_index 4

Avocado!

Huzzah!  You should now be ready to start recognizing your hotword requests.  Say “Avocado” a couple of times.  The should start coming out at the end of the trace like this (detected keyword):

Onward!  Now that you can be heard…

There are lots of things you can do.  For the lazy – simply run this process from your other programs… monitor stdout for that phrase and Voila!  You know when your software has been summoned!

Look for another article on using this in your NodeJS application… I just have to figure that out first… 😉

Some other articles you might like…

Wizard’s Fire!!!

[FIXED!!!] Turn on a lamp with a gesture – Image Processing! Machine learning!

Lessons on PI 6: Playing Audio on a Raspberry PI

The complete guide to enabling speech recognition on an RPI3… in NodeJS