adoxography/streamlined-genre

Masters project for the University of Washington's Master of Science in Computational Linguistics program. Concerns genre classification of audio recordings for low resource languages. # streamlined-genre ## Installation 0. Ensure that `python` (>=3.6), Java, and `openSMILE` are installed. `python3`, `java`, and `SMILExtract` should be available on the `PATH`. 1. Ensure that `pip`, `wheel`, and `setuptools` are installed and up-to-date ```pip install --upgrade pip wheel setuptools``` 2. Clone the repository and `cd` into it ```git clone github.com && cd streamlined-genre``` 3. Install the `streamlined-genre` package ```pip install .``` If you're developing `streamlined-genre`, you might want to install it with the optional dev dependencies instead: ```pip install -e ".[dev]"``` ## Usage ### Data preparation #### ELAR `streamlined-genre` is designed to work with directories generated from Endangered Languages Archive (ELAR) content. The directories should be constructed as follows: ``` {LANGUAGE} ├── Bundles │ └── {RECORDING_TITLE} │ └── {AUDIO_FILE} └── {LANGUAGE}_ELAR_Directory.csv ``` `{LANGUAGE}_ELAR_Directory.csv` should be structured as follows, where `{LABEL}` is the instance's genre label. Audio files may be in `wav` or `sph` format. ```csv ... {RECORDING_TITLE},...,{AUDIO_FILE},...{LABEL} ... ``` #### Non-ELAR `streamlined-genre` can also work with arbitrary `wav` files. The files should be in the same folder and be named `{IDENTIFIER}__{LABEL}.wav` (where `{IDENTIFIER}` is unique across all samples). ### Command line executable Once installed, the executable `streamlined-genre` is available on the command line. It requires at least one action flag: Flag | Action | Requires ---- | ------------------------------------------- | ------------------------------------ `-t` | Transfers audio files from ELAR directories | `--wavs` and at least one `--source` `-e` | Extracts LLDs and labels from wav files | `--wavs` and `--compiled` `-x` | Compiles bags of words from LLD files | `--compiled` `-c` | Classifies bags of words …