Benchmark for Nyumaya V1.X Make sure to use Branch V2.0 for 2.X Versions
Made in Vancouver, Canada by Picovoice
The purpose of this benchmarking framework is to provide a scientific comparison between different wake word detection engines in terms of accuracy and runtime metrics. While working on Porcupine we noted that there is a need for such a tool to empower customers to make data-driven decisions.
LibriSpeech (test_clean portion) is used as background dataset. It can be downloaded from OpenSLR.
Furthermore, more than 300 recordings of six keywords (alexa, computer, jarvis, smart mirror, snowboy, and view glass) from more than 50 distinct speakers are used. The recordings are crowd-sourced. The recordings are stored within the repository here.
In order to simulate real-world situations, the data is mixed with noise (at 10 dB SNR). For this purpose, we use DEMAND dataset which has noise recording in 18 different environments (e.g. kitchen, office, traffic, etc.). It can be downloaded from Kaggle.
Three wake-word engines are used. PocketSphinx which can be installed using PyPI. Porcupine and Snowboy which are included as submodules in this repository. The Snowboy engine has a audio frontend component which is not normally a part of wake word engines and is considered a separate part of audio processing chain. The other two engines have not such component in them. We enabled this component in Snowboy engine for this benchmark as this is the optimal way of running it.
We measure the accuracy of the wake word engines using false alarm per hour and miss detection rates. The false alarm per hour is measured as a number of false positives in an hour. Miss detection is measured as the percentage of wake word utterances an engine rejects incorrectly. Using these definitions we compare the engines for a given false alarm rate and the engine with a smaller miss detection rate has a better performance.
The measured runtime metric is real time factor. Real time factor is computed by dividing the processing time to the length of input audio. It can be thought of as average CPU usage. The engine with a lower real time factor is more computationally efficient (faster).
The benchmark has been developed on Ubuntu 18.04 with Python 3.6. Clone the repository using
git clone --recurse-submodules [email protected]:Picovoice/wakeword-benchmark.gitMake sure the Python packages in the requirements.txt are properly installed for your Python version as Python bindings are used for running the engines. The repositories for Porcupine and Snowboy are cloned in engines. Follow the instructions on their repositories to be able to run their Python demo before proceeding to the next step.
Usage information can be retrieved via
python benchmark.py -hThe benchmark can be run using the following command from the root of the repository
python benchmark.py --librispeech_dataset_path ${LIBRISPEECH_DATASET_PATH} --demand_dataset_path ${DEMAND_DATASET_PATH} --keyword ${KEYWORD}Refer to runtime documentation.
Below is the result of running the benchmark framework averaged on six different keywords. The plot below shows the miss rate of different engines at 1 false alarm per 10 hours. The lower the miss rate the more accurate the engine is.
Below is the runtime measurements on a Raspberry Pi 3. For Snowboy the runtime highly-depends on the keyword. Therefore we measured the CPU usage for each keyword and used the average.
