This is the repo for the SPADE AudioBNC cleaning script which makes a subset of high quality utterances from the corpus, split into speaker tiers.

Running the pipeline

To reproduce the dataset, all that's necessary is placing the requested textgrids in the input directory. You must have a symbolic link(or just a directory) to both the wav and textgrids directories, labeled wavs and textgrid respectively. It is also necessary to change the directory names at the top of the following scripts:do_pipeline.sh, aligner_difference.py, speaker_data.py. PREFIX should be changed to wherever you have put the pipeline folder, MFA_DIR should point to a directory containing the mfa_align binary for MFA. AUDIO_BNC_DIR should point to wherever the Texts directory of the BNC is located. Then, simply run the do_pipeline script, this will take a considerable amount of time(upwards of 24 hours) to run on the overall corpus.

Script description

1. output_dictionary.py: Runs over all textgrids and generates pronunciation.txt containing all words and their pronunciations for MFA.

2. output_mfa_formatted.py: Runs over all textgrids and replaces with labeled utterances for use in MFA. Also cuts each wav file to just the part used in a given textgrid again for MFA.

3. aligner-difference.py: Calculates HNR and aligner-difference for all textgrids in output

4. classify.py: Goes over textgrids outputted by aligner-differenc.py and decides whether to classify them as good or bad based on the previously described classifier.

5. speaker_data.py: Splits output from classify.py into speaker tiers based on the XML transcripts.

6. reduced_data_set.py: Goes over output from speaker_data.py and deletes all utterances not labeled "good". Additionally deletes tiers containing feature values.

Directory description

1. requirements.txt: List of required pip packages in python, to install run pip install -r requirements.txt

2. pronunciation.txt: List of pronunciations for all words in AudioBNC for MFA.

3. input: A directory with all the AudioBNC textgrids you wish to clean.

4. wavs: Directory or symlink to directory of all the AudioBNC wavs.

5. textgrid: Directory or symlink to directory of all the AudioBNC TextGrids.

6. output: Output from MFA

7. classify_grids: Output from aligner-difference.py, TextGrids which have yet to be classified.

8. corpus_for_mfa: Directory containing TextGrids to be used by MFA.

9. out_with_labels: Classified textgrids which have not yet been speakerised.

10. speakered_textgrids_chunked: Cleaned textgrids with labels describing quality of utterances, split into speaker tiers. Still contains all data, included feature-tiers.

11. cleaned_textgrids: Final product of pipeline, to be used in SPADE. Includes "good" utterances split into speaker-tiers.