How sanjaykishore18/agentslp_speech_pathologist works
AgentSLP is a research repository for training and evaluating two separate multi-label speech classifiers: a broad-problem model using text-derived and acoustic features, and SEP-28k stuttering-event models using annotated speech clips. The SEP workflow prepares clips from episode metadata and source audio, then trains and evaluates model variants; the repository also retains prior training scripts and generated model artifacts. The README documents the principal workflows, while sampled excerpts show preprocessing, model components, and evaluation details. Some artifact-writing and wiring details are unsampled, so those relationships are omitted or marked unverified.
Main entry points: SEP-28k_labels.csv SEP labels · prepare_sep28k.py Clip preparation · train_20k_model.py 20K classifier · train_sep28k_final_v3.py SEP v3 model
Components 16
Data inputs
- Broad features
- SEP labels
data/SEP-28k_labels.csv - Episode metadata
data/SEP-28k_episodes.csv
SEP preparation
- Clip preparation
prepare_sep28k.py - SEP clip corpus
Model training
- 20K classifier
train_20k_model.py - SEP classifier
train_sep28k_model.py - SEP v3 model
train_sep28k_final_v3.py - SEP WavLM model
train_sep28k_wavlm_final.py - Prior SEP model
backups/sep28k_model_previous_2026-09-24/train_sep28k_model.py - Prior corrected model
backups/sep28k_model_previous_2026-09-24/train_sep28k_model_final_corrected.py
Model artifacts
- 20K artifacts
artifacts_20k - SEP artifacts
artifacts_sep28k_final_v3 - WavLM artifacts
artifacts_sep28k_wavlm_final
Other components
- Researcher
- Episode source
Connections 21
- Researcher → 20K classifier: runs training
README.md - Researcher → Clip preparation: prepares audio
README.md - Researcher → SEP classifier: runs training
README.md - Broad features → 20K classifier: provides features
README.md - SEP labels → Clip preparation: provides annotations
prepare_sep28k.py - Episode metadata → Clip preparation: maps episode URLs
prepare_sep28k.py - Clip preparation → Episode source: downloads episodes
prepare_sep28k.py - Clip preparation → SEP clip corpus: writes clips
prepare_sep28k.py - SEP labels → SEP classifier: provides labels
README.md - SEP clip corpus → SEP classifier: provides audio
README.md - SEP labels → SEP v3 model: provides labels
train_sep28k_final_v3.py - SEP clip corpus → SEP v3 model: provides clips
train_sep28k_final_v3.py - SEP labels → SEP WavLM model: provides labels
train_sep28k_wavlm_final.py - SEP clip corpus → SEP WavLM model: provides clips
train_sep28k_wavlm_final.py - SEP labels → Prior SEP model: provides labels
backups/sep28k_model_previous_2026-09-24/train_sep28k_model.py - SEP clip corpus → Prior SEP model: provides clips
backups/sep28k_model_previous_2026-09-24/train_sep28k_model.py - SEP labels → Prior corrected model: provides labels
backups/sep28k_model_previous_2026-09-24/train_sep28k_model_final_corrected.py - SEP clip corpus → Prior corrected model: provides clips
backups/sep28k_model_previous_2026-09-24/train_sep28k_model_final_corrected.py - 20K classifier → 20K artifacts: saves results
- SEP v3 model → SEP artifacts: saves results
- SEP WavLM model → WavLM artifacts: saves results