Multilingual Acoustic Word Embeddings on GlobalPhone
Overview
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.
Disclaimer
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.
Download datasets
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.
Install dependencies
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
Extract speech features
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
Evaluate frame-level features using the same-different task
This is optional. To perform frame-level same-different evaluation based on dynamic time warping (DTW), follow samediff/readme.md.
Obtain downsampled acoustic word embeddings
Extract and evaluate downsampled acoustic word embeddings by running the steps in downsample/readme.md.
Train neural acoustic word embeddings
Train and evaluate neural network acoustic word embedding models by running the steps in embeddings/readme.md.
Analyse embedding models
Analyse different properties/aspects of the acoustic word embedding models by running the steps in blackbox/readme.md.
Query-by-example search
Perform query-by-example search experiments by running the steps in qbe/readme.md.
Unit tests
In the root project directory, run make test to run unit tests.
References
Contributors
License
The code is distributed under the Creative Commons Attribution-ShareAlike license (CC BY-SA 4.0).