ConnectedSpeech-UCSF

Data processing and analysis of connected speech features in PPA

Installation

1. Clone the repository

git clone -b main https://github.com/lingualab/ConnectedSpeech-UCSF
cd ConnectedSpeech-UCSF

2. Create a virtual environment

python3 -m venv --prompt ConnectedSpeech-UCSF venv
source venv/bin/activate

3. Install ConnectedSpeech-UCSF:

pip install -e .

Tasks description

Participants completed the Picnic Scene from the Western Aphasia Battery. Five transcripts are available:

• salt.slt: Original SALT transcript.
• salt.txt: SALT transcript with all SALT codes removed.
• manual.txt: SALT transcript with all SALT codes removed, and new lines removed.
• whisper.txt: Original Whiser transcript.
• whisperQC.txt: Whisper transcript manually checked, and disfluencies added.

Notes

• Updated data was sent from UCSF. We now use the data in /data/brambati/dataset/ConnectedSpeech-UCSF/sourcedata/NEW.

Scripts

• csucsf_process_merge: merge select speech features output from speechmetryflow with phenotype information for further analysis.
• csucsf_describe_participants: calculate demographic statistics based on data from included participants
• csucsf_analysis_wer: performs text analysis and compute Word Error Rates metric
• csucsf_analysis_icc: compute intraclass correlation analysis
• csucsf_classification: run binary classification for specified pairs of diagnoses, using linguistic features selected in the csucsf_process_merge script. Optionally include feature selection.
• csucsf_classification_stats: plot bar graphs comparing the classification performance of different transcription methods.