resproj007/pathological_speech

Pathological Speech (TORGO + UA-Speech + LibriSpeech Normal)

Mixed-corpus speech dataset for training and evaluating controllable speech-synthesis and severity-classification models. Three corpora are merged with unified metadata so a single model can learn severity- and gender-conditioned generation without confounds.

Splits (speaker-disjoint since 2026-09-14)

No speaker appears in both splits. The held-out speakers are TORGO_F03, TORGO_FC01, TORGO_M04, TORGO_MC02, UA-Speech_M09, LibriSpeech_1263, LibriSpeech_3242, LibriSpeech_6181, LibriSpeech_6818. test is the complete held-out set; evaluations that need fewer clips subsample it with a recorded seed rather than the dataset carrying a second split.

What was wrong before

Revision 71f6861aa9ea5bcc89d35fd62d8581853c54a43c claimed a held-out test set, but its test split was a per-clip sample of the SAME speakers as train: 17 of its 21 test speakers had training clips, and all 1500 TORGO and UA-Speech test rows came from speakers with 38,741 training rows. Only the 280 LibriSpeech test rows were speaker-disjoint. Every held-out number measured on that split was measured on voices the model had trained on. Load revision="71f6861aa9ea5bcc89d35fd62d8581853c54a43c" to reproduce that split. (Revision c76c0c09f329724734e7e1b93dd13dae4db69aa8 was the first speaker-disjoint rebuild; it split the held-out clips into test and heldout, now folded into one test.)

How the held-out speakers were chosen

Per gender x pathology cell over the two dysarthric corpora, one whole speaker per corpus is held out (a corpus with a single speaker in a cell cannot be held out: UA-Speech has one female speaker, UA-Speech_F04, so that corpus x cell is untestable). The joint assignment is a deterministic search that (1) keeps every severity band of every cell in train with at least one speaker and keeps every corpus x cell in train, (2) maximises the number of severity bands present in test, (3) minimises the largest share of any band's rows that leaves training, (4) minimises the total rows held out, ties broken by speaker id. LibriSpeech's four test speakers were already disjoint and are kept. 1 exact duplicate row(s) (same speaker, same audio bytes) were dropped.

Text is NOT disjoint: TORGO and UA-Speech use fixed prompt lists, so a test speaker's words and sentences are also spoken by training speakers.

Composition

By corpus:

By corpus, gender and severity:

By speaker:

Clips with four or more words: train 7888, test 1031.

Corpora

Why LibriSpeech?

Without it, every severity == Normal row came from TORGO controls, meaning a model could trivially learn Normal <-> TORGO acoustics instead of true prosodic normality. The LibriSpeech train-clean-100 supplement breaks that corpus x severity confound. All LibriSpeech rows are labelled severity="Normal", condition="Control", diagnosis="None".

Metadata canonicalization (applied to all shards)

1. 1.`gender` is always lowercase ("male" / "female").
2. 2.`speaker_id` is prefixed with `{corpus}_` (e.g. TORGO_M01, UA-Speech_M01, LibriSpeech_3242). TORGO and UA-Speech both use M01 / M05 for different humans.
3. 3.`severity` is recomputed per speaker from the canonical mapping (TORGO: clinical severity from the TORGO speaker sheet; UA-Speech: intelligibility banding -> Very-low (<=19%) -> Severe, Low (28-39%) / Mid (58-62%) -> Moderate, High (>=86%) -> Mild).

Other columns (diagnosis, condition, intelligibility) are preserved verbatim from their source corpora. Audio bytes are unchanged from the original revision, so content-addressed keys (sha1 of the encoded bytes) still match.

Schema

Load

from datasets import load_dataset
ds = load_dataset("resproj007/pathological_speech")
print(ds)
# DatasetDict({
#   train: Dataset({ features: [...], num_rows: 37704 })
#   test:  Dataset({ features: [...], num_rows: 8812 })
# })

Intended use

• —Controllable pathological-speech synthesis (severity + gender conditioning).
• —Severity classification with leave-speaker-out evaluation.
• —Cross-corpus robustness analysis (TORGO <-> UA-Speech).

Licensing

Derived from corpora released under research-use terms. Redistribution is intended for academic research; verify the TORGO, UA-Speech, and LibriSpeech licenses for your use case.