Multilingual acoustic word embedding approaches are implemented and evaluated on the GlobalPhone corpus. The experiments are described in:
- H. Kamper, Y. Matusevych, and S.J. Goldwater, "Multilingual acoustic word embedding models for processing zero-resource languages," in Proc. ICASSP, 2020. [arXiv]
- H. Kamper, Y. Matusevych, and S. J. Goldwater, "Improved acoustic word embeddings for zero-resource languages using multilingual transfer," arXiv preprint arXiv:2006.02295, 2020. [arXiv]
Please cite these papers if you use the code.
The code provided here is not pretty. But I believe that research should be reproducible. I provide no guarantees with the code, but please let me know if you have any problems, find bugs or have general comments.
The GlobalPhone
corpus and forced alignments of the data needs to be obtained. GlobalPhone
needs to be paid for. If you have proof of payment, we can give you access to
the forced alignments. Save the data and forced alignments in a separate
directory and update the paths.py file to point to the data directories.
You will require the following:
To install speech_dtw (required for same-different evaluation) and shorten
(required for processing audio), run ./install_local.sh.
You can install all the other dependencies in a conda environment by running:
conda env create -f environment.yml
conda activate tf1.13
Update the paths in paths.py to point to the data directories. Extract MFCC
features in the features/ directory as follows:
cd features
./extract_features.py SP
You need to run extract_features.py for all languages; run it without any
arguments to see all 16 language codes.
UTD pairs can also be analysed here, by running e.g.:
./analyse_utd_pairs.py SP
This is optional. To perform frame-level same-different evaluation based on dynamic time warping (DTW), follow samediff/readme.md.
Extract and evaluate downsampled acoustic word embeddings by running the steps in downsample/readme.md.
Train and evaluate neural network acoustic word embedding models by running the steps in embeddings/readme.md.
Analyse different properties/aspects of the acoustic word embedding models by running the steps in blackbox/readme.md.
Perform query-by-example search experiments by running the steps in qbe/readme.md.
In the root project directory, run make test to run unit tests.
The code is distributed under the Creative Commons Attribution-ShareAlike license (CC BY-SA 4.0).