Honk: CNNs for Keyword Spotting

Honk is a PyTorch reimplementation of Google's TensorFlow convolutional neural networks for keyword spotting, which accompanies the recent release of their Speech Commands Dataset. For more details, please consult our writeup:

• Raphael Tang, Jimmy Lin. Honk: A PyTorch Reimplementation of Convolutional Neural Networks for Keyword Spotting. arXiv:1710.06554, October 2017.

Honk is useful for building on-device speech recognition capabilities for interactive intelligent agents. Our code can be used to identify simple commands (e.g., "stop" and "go") and be adapted to detect custom "command triggers" (e.g., "Hey Siri!").

Check out this video for a demo of Honk in action!

Demo Application

Use the instructions below to run the demo application (shown in the above video) yourself!

Currently, PyTorch has official support for only Linux and OS X. Thus, Windows users will not be able to run this demo easily.

To deploy the demo, run the following commands:

• If you do not have PyTorch, please see the website.
• Install Python dependencies: pip install -r requirements.txt
• Install GLUT (OpenGL Utility Toolkit) through your package manager (e.g. apt-get install freeglut3-dev)
• Fetch the data and models: ./fetch_data.sh
• Start the PyTorch server: python .
• Run the demo: python utils/speech_demo.py

If you need to adjust options, like turning off CUDA, please edit config.json.

Additional notes for Mac OS X:

• GLUT is already installed on Mac OS X, so that step isn't needed.
• If you have issues installing pyaudio, this may be the issue.

Server

Setup and deployment

python . deploys the web service for identifying if audio contain the command word. By default, config.json is used for configuration, but that can be changed with --config=<file_name>. If the server is behind a firewall, one workflow is to create an SSH tunnel and use port forwarding with the port specified in config (default 16888).

In our honk-models repository, there are several pre-trained models for Caffe2 (ONNX) and PyTorch. The fetch_data.sh script fetches these models and extracts them to the model directory. You may specify which model and backend to use in the config file's model_path and backend, respectively. Specifically, backend can be either caffe2 or pytorch, depending on what format model_path is in. Note that, in order to run our ONNX models, the packages onnx and onnx_caffe2 must be present on your system; these are absent in requirements.txt.

Raspberry Pi (RPi) Infrastructure Setup

Unfortunately, getting the libraries to work on the RPi, especially librosa, isn't as straightforward as running a few commands. We outline our process, which may or may not work for you.

1. Obtain an RPi, preferably an RPi 3 Model B running Raspbian. Specifically, we used this version of Raspbian Stretch.
2. Install dependencies: sudo apt-get install -y protobuf-compiler libprotoc-dev python-numpy python-pyaudio python-scipy python-sklearn
3. Install Protobuf: pip install protobuf
4. Install ONNX without dependencies: pip install --no-deps onnx
5. Follow the official instructions for installing Caffe2 on Raspbian. This process takes about two hours. You may need to add the caffe2 module path to the PYTHONPATH environment variable. For us, this was accomplished by export PYTHONPATH=$PYTHONPATH:/home/pi/caffe2/build
6. Install the ONNX extension for Caffe2: pip install onnx-caffe2
7. Install further requirements: pip install -r requirements_rpi.txt
8. Install librosa: pip install --no-deps resampy librosa
9. Try importing librosa: python -c "import librosa". It should throw an error regarding numba, since we haven't installed it.
10. We haven't found a way to easily install numba on the RPi, so we need to remove it from resampy. For our setup, we needed to remove numba and @numba.jit from /home/pi/.local/lib/python2.7/site-packages/resampy/interpn.py
11. All dependencies should now be installed. We should try deploying an ONNX model.
12. Fetch the models and data: ./fetch_data.sh
13. In config.json, change backend to caffe2 and model_path to model/google-speech-dataset-full.onnx.
14. Deploy the server: python . If there are no errors, you have successfully deployed the model, accessible via port 16888 by default.
15. Run the speech commands demo: python utils/speech_demo.py. You'll need a working microphone and speakers. If you're interacting with your RPi remotely, you can run the speech demo locally and specify the remote endpoint --server-endpoint=http://[RPi IP address]:16888.

Utilities

QA client

Unfortunately, the QA client has no support for the general public yet, since it requires a custom QA service. However, it can still be used to retarget the command keyword.

python client.py runs the QA client. You may retarget a keyword by doing python client.py --mode=retarget. Please note that text-to-speech may not work well on Linux distros; in this case, please supply IBM Watson credentials via --watson-username and --watson--password. You can view all the options by doing python client.py -h.

Training and evaluating the model

CNN models. python -m utils.train --mode [train|eval] trains or evaluates the model. It expects all training examples to follow the same format as that of Speech Commands Dataset. The recommended workflow is to download the dataset and add custom keywords, since the dataset already contains many useful audio samples and background noise.

Residual models. We recommend the following hyperparameters for training any of our res{8,15,26}[-narrow] models on the Speech Commands Dataset:

python -m utils.train --wanted_words yes no up down left right on off stop go --dev_every 1 --n_labels 12 --n_epochs 26 --weight_decay 0.00001 --lr 0.1 0.01 0.001 --schedule 3000 6000 --model res{8,15,26}[-narrow]

For more information about our deep residual models, please see our paper:

• Raphael Tang, Jimmy Lin. Deep Residual Learning for Small-Footprint Keyword Spotting. Proceedings of the 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2018), April 2018, Calgary, Alberta, Canada.

There are command options available:

option	input format	default	description
--batch_size	[1, n)	100	the mini-batch size to use
--cache_size	[0, inf)	32768	number of items in audio cache, consumes around 32 KB * n
--conv1_pool	[1, inf) [1, inf)	2 2	the width and height of the pool filter
--conv1_size	[1, inf) [1, inf)	10 4	the width and height of the conv filter
--conv1_stride	[1, inf) [1, inf)	1 1	the width and length of the stride
--conv2_pool	[1, inf) [1, inf)	1 1	the width and height of the pool filter
--conv2_size	[1, inf) [1, inf)	10 4	the width and height of the conv filter
--conv2_stride	[1, inf) [1, inf)	1 1	the width and length of the stride
--data_folder	string	/data/speech_dataset	path to data
--dev_every	[1, inf)	10	dev interval in terms of epochs
--dev_pct	[0, 100]	10	percentage of total set to use for dev
--dropout_prob	[0.0, 1.0)	0.5	the dropout rate to use
--gpu_no	[-1, n]	1	the gpu to use
--group_speakers_by_id	{true, false}	true	whether to group speakers across train/dev/test
--input_file	string		the path to the model to load
--input_length	[1, inf)	16000	the length of the audio
--lr	(0.0, inf)	{0.1, 0.001}	the learning rate to use
--mode	{train, eval}	train	the mode to use
--model	string	cnn-trad-pool2	one of cnn-trad-pool2, cnn-tstride-{2,4,8}, cnn-tpool{2,3}, cnn-one-fpool3, cnn-one-fstride{4,8}, res{8,15,26}[-narrow], cnn-trad-fpool3, cnn-one-stride1
--momentum	[0.0, 1.0)	0.9	the momentum to use for SGD
--n_dct_filters	[1, inf)	40	the number of DCT bases to use
--n_epochs	[0, inf)	500	number of epochs
--n_feature_maps	[1, inf)	{19, 45}	the number of feature maps to use for the residual architecture
--n_feature_maps1	[1, inf)	64	the number of feature maps for conv net 1
--n_feature_maps2	[1, inf)	64	the number of feature maps for conv net 2
--n_labels	[1, n)	4	the number of labels to use
--n_layers	[1, inf)	{6, 13, 24}	the number of convolution layers for the residual architecture
--n_mels	[1, inf)	40	the number of Mel filters to use
--no_cuda	switch	false	whether to use CUDA
--noise_prob	[0.0, 1.0]	0.8	the probability of mixing with noise
--output_file	string	model/google-speech-dataset.pt	the file to save the model to
--seed	(inf, inf)	0	the seed to use
--silence_prob	[0.0, 1.0]	0.1	the probability of picking silence
--test_pct	[0, 100]	10	percentage of total set to use for testing
--timeshift_ms	[0, inf)	100	time in milliseconds to shift the audio randomly
--train_pct	[0, 100]	80	percentage of total set to use for training
--unknown_prob	[0.0, 1.0]	0.1	the probability of picking an unknown word
--wanted_words	string1 string2 ... stringn	command random	the desired target words

Recording audio

You may do the following to record sequential audio and save to the same format as that of speech command dataset:

python -m utils.record

Input return to record, up arrow to undo, and "q" to finish. After one second of silence, recording automatically halts.

Several options are available:

--output-begin-index: Starting sequence number
--output-prefix: Prefix of the output audio sequence
--post-process: How the audio samples should be post-processed. One or more of "trim" and "discard_true".

Post-processing consists of trimming or discarding "useless" audio. Trimming is self-explanatory: the audio recordings are trimmed to the loudest window of x milliseconds, specified by --cutoff-ms. Discarding "useless" audio (discard_true) uses a pre-trained model to determine which samples are confusing, discarding correctly labeled ones. The pre-trained model and correct label are defined by --config and --correct-label, respectively.

For example, consider python -m utils.record --post-process trim discard_true --correct-label no --config config.json. In this case, the utility records a sequence of speech snippets, trims them to one second, and finally discards those not labeled "no" by the model in config.json.

Listening to sound level

python manage_audio.py listen

This assists in setting sane values for --min-sound-lvl for recording.

Generating contrastive examples

python manage_audio.py generate-contrastive --directory [directory] generates contrastive examples from all .wav files in [directory] using phonetic segmentation.

Trimming audio

Speech command dataset contains one-second-long snippets of audio.

python manage_audio.py trim --directory [directory] trims to the loudest one-second for all .wav files in [directory]. The careful user should manually check all audio samples using an audio editor like Audacity.