pvporcupinedemo 1.9.7

Porcupine wake word engine demos

Porcupine Wake Word Engine Demos

Made in Vancouver, Canada by Picovoice

This package contains demos and commandline utilities for processing real-time audio (i.e. microphone) and audio files using Porcupine wake word engine.

Porcupine

Porcupine is a highly-accurate and lightweight wake word engine. It enables building always-listening voice-enabled applications. It is

• using deep neural networks trained in real-world environments.
• compact and computationally-efficient making it perfect for IoT.
• scalable. It can detect multiple always-listening voice commands with no added CPU/memory footprint.
• self-service. Developers can train custom wake phrases using Picovoice Console.

Compatibility

• Python 3
• Runs on Linux (x86_64), Mac (x86_64), Windows (x86_64), Raspberry Pi (all variants), NVIDIA Jetson (Nano), and BeagleBone.

Installation

sudo pip3 install pvporcupinedemo

Usage

Microphone Demo

It opens an audio stream from a microphone and detects utterances of a given wake word. The following opens the default microphone and detects occurrences of "Picovoice".

porcupine_demo_mic --keywords picovoice

keywords is a shorthand for using default keyword files shipped with the package. The list of default keyword files can be seen in the usage string

porcupine_demo_mic --help

To detect multiple phrases concurrently provide them as separate arguments

porcupine_demo_mic --keywords picovoice porcupine

To detect non-default keywords (e.g. models created using Picovoice Console) use keyword_paths argument

porcupine_demo_mic --keyword_paths ${KEYWORD_PATH_ONE} ${KEYWORD_PATH_TWO}

To detect non-English keywords provide the respective model path:

porcupine_demo_mic --model_path ${NON_ENGLISH_MODEL_PATH} --keyword_paths ${NON_ENGLISH_KEYWORD_PATH} 

The model files for all supported languages are available here on Porcupine's GitHub repository.

It is possible that the default audio input device recognized by PyAudio is not the one being used. There are a couple of debugging facilities baked into the demo application to solve this. First, type the following into the console

porcupine_demo_mic --show_audio_devices

It provides information about various audio input devices on the box. On a Linux box, this is the console output

index: 0, device name: Monitor of sof-hda-dsp HDMI3/DP3 Output
index: 1, device name: Monitor of sof-hda-dsp HDMI2/DP2 Output
index: 2, device name: Monitor of sof-hda-dsp HDMI1/DP1 Output
index: 3, device name: Monitor of sof-hda-dsp Speaker + Headphones
index: 4, device name: sof-hda-dsp Headset Mono Microphone + Headphones Stereo Microphone
index: 5, device name: sof-hda-dsp Digital Microphone

If you would like to use the default device, leave audio_device_index empty, else select the device index from the output above. In this example we will use the device at index 5.

porcupine_demo_mic --keywords picovoice --audio_device_index 5

If the problem persists we suggest storing the recorded audio into a file for inspection. This can be achieved by

porcupine_demo_mic --keywords picovoice --audio_device_index 5 --output_path ~/test.wav

If after listening to stored file there is no apparent problem detected please open an issue.

File Demo

It allows testing Porcupine on a corpus of audio files. The demo is mainly useful for quantitative performance benchmarking. It accepts 16kHz audio files. Porcupine processes a single-channel audio stream if a stereo file is provided it only processes the first (left) channel. The following processes a file looking for instances of the phrase "Picovoice"

porcupine_demo_file --input_audio_path ${AUDIO_PATH} --keywords picovoice

keywords is a shorthand for using default keyword files shipped with the package. The list of default keyword files can be seen in the usage string

porcupine_demo_file --help

To detect multiple phrases concurrently provide them as separate arguments

porcupine_demo_file --input_audio_path ${AUDIO_PATH} --keywords grasshopper porcupine

To detect non-default keywords (e.g. models created using Picovoice Console) use keyword_paths argument

porcupine_demo_file --input_audio_path ${AUDIO_PATH} \
--keyword_paths ${KEYWORD_PATH_ONE} ${KEYWORD_PATH_TWO}

To detect non-English keywords provide the respective model path:

porcupine_demo_mic --input_audio_path ${AUDIO_PATH} \
--model_path ${NON_ENGLISH_MODEL_PATH} \
--keyword_paths ${NON_ENGLISH_KEYWORD_PATH} 

The model files for all supported languages are available here on Porcupine's GitHub repository.

The sensitivity of the engine can be tuned per keyword using the sensitivities input argument

porcupine_demo_file --input_audio_path ${AUDIO_PATH} \
--keywords grasshopper porcupine --sensitivities 0.3 0.6

Sensitivity is the parameter that enables trading miss rate for the false alarm rate. It is a floating point number within [0, 1]. A higher sensitivity reduces the miss rate at the cost of increased false alarm rate.