Wake word detection modeling for Firefox Voice, supporting open datasets like Google Speech Commands and Mozilla Common Voice.
Citation:
@inproceedings{tang-etal-2020-howl,
title = "Howl: A Deployed, Open-Source Wake Word Detection System",
author = "Tang, Raphael and Lee, Jaejun and Razi, Afsaneh and Cambre, Julia and Bicking, Ian and Kaye, Jofish and Lin, Jimmy",
booktitle = "Proceedings of Second Workshop for NLP Open Source Software (NLP-OSS)",
month = nov,
year = "2020",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2020.nlposs-1.9",
doi = "10.18653/v1/2020.nlposs-1.9",
pages = "61--65"
}
A proper Pip package is coming soon.
-
git clone https://github.com/castorini/howl && cd howl
-
Install PyTorch by following your platform-specific instructions.
-
Install PyAudio and its dependencies through your distribution's package system.
-
pip install -r requirements.txt
(some apt packages might need to be installed)
In the example that follows, we describe how to train a custom detector for the word, "fire."
- Download a supported data source. We recommend Common Voice for its breadth and free license.
- To provide alignment for the data, install Montreal Forced Aligner (MFA) and download an English pronunciation dictionary.
- Create a positive dataset containing the keyword:
VOCAB='["fire"]' INFERENCE_SEQUENCE=[0] DATASET_PATH=data/fire-positive python -m howl.run.create_raw_dataset --negative-pct 0 -i ~/path/to/common-voice --positive-pct 100
- Create a negative dataset without the keyword:
VOCAB='["fire"]' INFERENCE_SEQUENCE=[0] DATASET_PATH=data/fire-negative python -m howl.run.create_raw_dataset --negative-pct 5 -i ~/path/to/common-voice --positive-pct 0
- Generate some mock alignment for the negative set, where we don't care about alignment:
DATASET_PATH=data/fire-negative python -m howl.run.attach_alignment --align-type stub
- Use MFA to generate alignment for the positive set:
mfa_align data/fire-positive/audio eng.dict pretrained_models/english.zip output-folder
- Attach the MFA alignment to the positive dataset:
DATASET_PATH=data/fire-positive python -m howl.run.attach_alignment --align-type mfa -i output-folder
- Source the relevant environment variables for training the
res8
model:source envs/res8.env
. - Train the model:
python -m howl.run.train -i data/fire-negative data/fire-positive --model res8 --workspace workspaces/fire-res8
. - For the CLI demo, run
python -m howl.run.demo --model res8 --workspace workspaces/fire-res8
.
howl-models contains workspaces with pretrained models
to get the latest models, simply run git submodule update --init --recursive
VOCAB='[" hey","fire","fox"]' INFERENCE_SEQUENCE=[0,1,2] INFERENCE_THRESHOLD=0 NUM_MELS=40 MAX_WINDOW_SIZE_SECONDS=0.5 python -m howl.run.demo --model res8 --workspace howl-models/howl/hey-fire-fox
First, follow the installation instructions in the quickstart guide.
- Download the Google Speech Commands dataset and extract it.
- Source the appropriate environment variables:
source envs/res8.env
- Set the dataset path to the root folder of the Speech Commands dataset:
export DATASET_PATH=/path/to/dataset
- Train the
res8
model:NUM_EPOCHS=20 MAX_WINDOW_SIZE_SECONDS=1 VOCAB='["yes","no","up","down","left","right","on","off","stop","go"]' BATCH_SIZE=64 LR_DECAY=0.8 LEARNING_RATE=0.01 python -m howl.run.pretrain_gsc --model res8
- Download the Hey Firefox corpus, licensed under CC0, and extract it.
- Download our noise dataset, built from Microsoft SNSD and MUSAN, and extract it.
- Source the appropriate environment variables:
source envs/res8.env
- Set the noise dataset path to the root folder:
export NOISE_DATASET_PATH=/path/to/snsd
- Set the firefox dataset path to the root folder:
export DATASET_PATH=/path/to/hey_firefox
- Train the model:
LR_DECAY=0.98 VOCAB='[" hey","fire","fox"]' USE_NOISE_DATASET=True BATCH_SIZE=16 INFERENCE_THRESHOLD=0 NUM_EPOCHS=300 NUM_MELS=40 INFERENCE_SEQUENCE=[0,1,2] MAX_WINDOW_SIZE_SECONDS=0.5 python -m howl.run.train --model res8 --workspace workspaces/hey-ff-res8
- Download hey snips dataset
- Process the dataset to a format howl can load
VOCAB='["hey","snips"]' INFERENCE_SEQUENCE=[0,1] DATASET_PATH=data/hey-snips python -m howl.run.create_raw_dataset --dataset-type 'hey-snips' -i ~/path/to/hey_snips_dataset
- Generate some mock alignment for the dataset, where we don't care about alignment:
DATASET_PATH=data/hey-snips python -m howl.run.attach_alignment --align-type stub
- Use MFA to generate alignment for the dataset set:
mfa_align data/hey-snips/audio eng.dict pretrained_models/english.zip output-folder
- Attach the MFA alignment to the dataset:
DATASET_PATH=data/hey-snips python -m howl.run.attach_alignment --align-type mfa -i output-folder
- Source the appropriate environment variables:
source envs/res8.env
- Set the noise dataset path to the root folder:
export NOISE_DATASET_PATH=/path/to/snsd
- Set the noise dataset path to the root folder:
export DATASET_PATH=/path/to/hey-snips
- Train the model:
LR_DECAY=0.98 VOCAB='[" hey","snips"]' USE_NOISE_DATASET=True BATCH_SIZE=16 INFERENCE_THRESHOLD=0 NUM_EPOCHS=300 NUM_MELS=40 INFERENCE_SEQUENCE=[0,1] MAX_WINDOW_SIZE_SECONDS=0.5 python -m howl.run.train --model res8 --workspace workspaces/hey-snips-res8
To verify the correctness of our implementation, we first train and evaluate our models on the Google Speech Commands dataset, for which there exists many known results. Next, we curate a wake word detection datasets and report our resulting model quality.
For both experiments, we generate reports in excel format. experiments folder includes sample outputs from the for each experiment and corresponding workspaces can be found here
For command recognition, we train the four different models (res8, LSTM, LAS encoder, MobileNetv2) to detect twelve different keywords: “yes”, “no”, “up”, “down”, “left”, “right”, “on”, “off”, “stop”, “go”, unknown, or silence.
python -m howl.run.eval_commands_recognition --num_iterations n --dataset_path < path_to_gsc_datasets >
In this experiment, we train our best commands recognition model, res8, for hey firefox
and hey snips
and evaluate them with different threashold.
Two different performance reports are generated, one with the clean audio and one with audios with noise
python -m howl.run.eval_wake_word_detection --num_models n --hop_size < number between 0 and 1 > --exp_type < hey_firefox | hey_snips > --dataset_path "x" --noiseset_path "y"
We also provide a script for generating ROC curve. exp_timestamp
can be found from the reports generated from previous command
python -m howl.run.generate_roc --exp_timestamp < experiment timestamp > --exp_type < hey_firefox | hey_snips >