sl-cedar1

⚠️ PRIVATE — DO NOT SHARE

This audio dataset is private and cannot be shared with anyone.

It contains customer voice recordings and the voice of a named individual, held under a private commercial engagement. It must not be redistributed, mirrored, published, quoted, used in a public demo, or included in any released model card, paper artifact, or dataset index. Access is limited to the repository owner. If you are reading this and you are not the owner, you should not have this file — report it and delete your copy.

Saidzoda Lab · private voice corpora · assembled 2026-09-02.

---

Contents

proj2_2voice/ — customer two-speaker corpus (ledger #14)

Synthetic read speech in two cloned customer voices, generated with Chatterbox-MTL-v3 (cb_step8000) from the customer's own document set.

Speakers	spk1, spk2 — real names deliberately never recorded
Generated clips	26,593 per speaker (53,186 total)
Text pool	26,593 unique normalized sentences from 994 customer documents (100% customer text, no lab top-up)
Source refs	15 s max-RMS windows cut from a 79.3 s (spk1) and 40.6 s (spk2) original recording
QC	round-trip WER via Whisper-large-v3-tg checkpoint-14450; clips above WER 0.15 were deleted at generation time (drop counts retained, audio not)
Voice split	80% cloned voice / 20% locked narrator reference

Layout: gen/spk{1,2}/wav/ (audio), gen/spk{1,2}/asr_manifest.s*.jsonl (transcripts + paths), refs/ (the two 15 s reference clips), raw/ (original mp3s + customer source texts), pool/ (sentence pool), code/ + run/ (generation scripts, MANIFEST, code hashes).

dilshod/ — named-speaker domain-adaptation corpus and evaluation

Clone-voice synthetic corpus plus the real-dictation evaluation pair that proved per-customer ASR domain adaptation (Whisper 21.5% → 12.1% real-note WER; Parakeet 25.1% → 22.7%).

domain_corpus/	1,563 clips, 3.54 h, 24 kHz mono 16-bit PCM, cloned from a 15 s reference
voicediag_eval/	akaiDilshod_16k.wav — real human dictation, plus 5 noise-condition variants and the ground-truth transcript
asr_splits/	train / val / test_synth jsonl

Caveat on asr_splits/train.jsonl: 5,903 rows, but only the 1,563 domain-corpus clips live here — the remaining ~4,340 are general-replay rows whose audio_filepath points outside this repo. Those paths are dangling in this copy by design; the replay audio is lab-general data and was deliberately not mixed into a private customer repo.

---

Provenance

Both corpora were produced on the ca-factory H100 box. Every run carries its own MANIFEST.yaml (code hash, pip freeze, resolved config) and, where the run is closed, an OUTCOME.md — those files are included so the corpora stay reproducible without reference to the box, which is ephemeral.

Publishing decision recorded in proj2_2voice/run/MANIFEST.yaml: "Final destination = a TRULY PRIVATE HF repo (not public, not gated); operator decision 2026-09-01 overriding zero-leak-to-HF." This repo is private, not gated — gated repos are still listable and requestable, which is not appropriate for this material.