midralab/gol-dala-cluster

GOL voice clusters — audited repair

This audit covers midralab/gol-dala-cluster revision 89e1c5982086207f0a8de22cdb880e5ea52789f6. The original repository has no dataset card and stores its files under Windows-style backslash paths. Its 89-byte cluster_statistics.json ends inside the cluster_statistics object and is invalid JSON.

Verified source layout

• voice_clusters.csv: 7,362,684 strictly parsed rows, 596 folders, 19,245 speakers, 150 cluster IDs (0–149), and 10,643.669 hours.
• voice_embeddings.npy: NumPy v1, little-endian float64, shape (7,362,684, 32). Its exact expected byte length, including the 128-byte header, equals the 1,884,847,232-byte repository object.
• cluster_statistics.json: truncated and invalid; replace it with the generated cluster_statistics.json.

The CSV row count and embedding first dimension match exactly. The NPY has no embedded row IDs, so the CSV order is part of the data contract and must not be changed without applying the identical permutation to the embedding matrix.

CSV audit

• Duration range: 0.5–29.9976 s; median 4.5745 s; p95 11.4337 s.
• Every file_exists value is true; no empty text or stored text_length mismatch.
• All full_audio_path values are non-portable D:\gal-dataset\... paths. Their basename, folder ID, and speaker ID are internally consistent.
• 148,170 rows contain no detected Japanese script. Most frequent cases are punctuation-only utterances and should be reviewed with audio.
• 217,477 repeated-text groups contribute 1,041,892 occurrences beyond the first.
• 541 folders have 42,364 gaps in their local file_index ranges. The composite (folder_id, file_index) itself has no duplicates.

Same-audio conflicts and conservative manifest

The same full_audio_path occurs in 8,911 groups, totaling 10,624 rows beyond the first. All duplicates agree on duration and voice cluster, but 2,370 groups (5,040 rows) attach more than one transcript to the same audio path. The largest observed group maps one WAV to 42 rows and 37 distinct transcripts. These are not safe to resolve without listening.

The audit's conservative manifest therefore:

1. excludes every row from a conflicting-transcript audio-path group;
2. collapses 6,541 non-conflicting repeated paths to the minimum file_index row;
3. removes the local Windows path and emits portable audio_relpath plus source_path_sha256;
4. keeps 7,349,690 conservative rows and a has_japanese_script flag.

The 5,040 excluded conflict rows are preserved in the audit's conflict table for audio-assisted adjudication. This is a reversible overlay; the source CSV is unchanged.

The repaired cluster_statistics.json and a machine-readable audit summary are published in this repository.

Rights and provenance

No license, title-level provenance, embedding model/version, clustering algorithm, distance metric, or random seed was present. Repository gating does not supply those terms. Before publication or training use, document the audio rights, the source GOL revision, the 32-dimensional embedding recipe, and how cluster IDs were produced.