instinct-org/all15_speaker_deduped_tts_train_clone_pairs_raw

All-15 Speaker-Deduped TTS Train Clone Pairs Raw

This dataset contains raw metadata rows for speaker-deduped TTS voice-clone training pairs. It does not contain audio bytes. Rows point back to source audio records and include reference/target metadata, language, dataset, tier, and precomputed speaker-similarity fields from the mining pipeline.

Contents

• data/train/distinct_speaker_clone_pair_plan.jsonl.gz: all survivor rows.
• data/by_dataset/*.jsonl.gz: the same survivor rows split by source dataset for filtering and audit.
• metadata/distinct_speaker_clone_pair_plan_summary.json: generation summary and counts.

The rejected duplicate-speaker rows are intentionally not included in this dataset. They are audit artifacts and should not be used as positive clone training rows for the baseline.

Counts

Metric	Count
Raw A/B input rows	181,843
Survivor rows	55,249
Rejected duplicate-speaker rows	126,594
Survivor Uzbek rows	5,887
Survivor Russian rows	49,362
Survivor A rows	1,404
Survivor B rows	53,845

Survivor Rows By Dataset

Dataset	Rows
yt4_chunked_speech_restorised	18,536
yt2_chunked_speech_restorised	9,742
yt_chunked_speech_restorised	7,953
yt3_chunked_speech_restorised	7,063
miscellaneous_yt_chunked_speech_restorised	5,805
tbp_chunked_speech_restorised	3,139
yt1_chunked_speech_restorised	2,925
audiobook_chunked_speech_restorised	77
default_voices_chunked_speech_restorised	5
espeech_podcasts_chunked_speech_restorised	4

Intended Use

Use these rows as a speaker-deduped clone-pair training pool with C-tier excluded. For balanced Uzbek/Russian training, Uzbek availability is the limiting side. Reserve a disjoint B-heavy clone heldout/dev set before selecting train rows.

Recommended baseline constraints:

train tiers: A and B only
heldout/dev tiers: B first, A only if needed
C tier: excluded
language balance: 50% Uzbek / 50% Russian
speaker proxy reuse: one survivor row per speaker-proxy component

Notes

The speaker count is a speaker-proxy count derived from the pair graph and source-seed merges, not a guaranteed count of unique real humans. The dedupe policy keeps one high-confidence survivor per speaker-proxy component to reduce speaker domination during training.