Picovoice Koala Noise Suppression Benchmark

Prerequisites

• Ubuntu 20.04 (x86_64)
• Git
• Python 3
• PIP
• Picovoice Console Account

Usage

1. Clone the repository:

git clone https://github.com/Picovoice/noise-suppression-benchmark.git
cd noise-suppression-benchmark

2. Install the dependencies:

pip3 install -r requirements.txt

3. Follow the instructions in README.md to obtain the Interspeech2020 dataset for the benchmark.

4. Log in to (or sign up for free) Picovoice Console and copy your AccessKey.

5. Run the benchmark:

python3 benchmark.py \
--engine picovoice_koala \
--dataset Interspeech2020 \
--data-folder ${PATH_TO_DNS_CHALLENGE_FOLDER}/datasets/test_set/synthetic/no_reverb \
--picovoice-access-key ${PICOVOICE_ACCESS_KEY}

Replace ${PATH_TO_DNS_CHALLENGE_FOLDER} with the path to the main folder of the DNS Challenge repository that you cloned to obtain the data, and ${PICOVOICE_ACCESS_KEY} with your AccessKey. Take note of the STOI metric printed to the terminal.

By default, the command uses the original mix of clean speech and noise as given by the dataset. To evaluate the performance at a specific noise level, the benchmark can also re-mix the data at a custom signal-to-noise ratio (SNR). Simply repeat the command above with an additional argument --remix-snr-db ${SNR_DB}, where ${SNR_DB} is replaced with the SNR in decibels (dB).