NagaYu/bleep-spans

STATUS: maintained

The model trained on this corpus, NagaYu/bleep-0.09b, is closed at its problem framing after one convergence retrain. This dataset is maintained: its files, splits, labels and licence stay as published.

Bleep spans — synthetic sensitive-speech regions with frame-accurate labels

Where sensitive information is spoken, and what kind it is — never what was said.

Synthetic only, English only. Every recording is synthetic. No real telephone call, clinical recording, or any other real speech was used, recorded, or derived from at any stage. Only English is included; Japanese generators exist in the code but are not represented here and are not benchmarked.

The most useful thing this corpus taught: its training conditions were too mild

If you build on this corpus, start here. The degradations that sound hard are not the ones that break a modern ASR on TTS speech.

On the eight training conditions, a whisper-tiny.en + text-PII baseline is essentially flat: 0.818 recall on clean audio and 0.857 at 8 kHz — better than clean. Band-limiting and G.711 μ-law are not hard conditions for a modern ASR on hyper-articulated TTS speech, and neither are 5–10 dB of babble or street noise, moderate reverberation or microphone colouring. The baseline only collapsed in the held-out set, and only where talkers overlap: 0.271 at 0 dB three-way overlap and 0.490 at narrowband-plus-overlap.

A second, harsher held-out set built later (test_heldout_v2, not part of these files — see below) confirms that pattern: the baseline collapses again only under overlap (0.196 at −3 dB), and holds 0.80–0.91 under a music bed at 5 dB, a low-bitrate codec, hard clipping, and AGC pumping with dropouts.

So: if you are using this corpus to study transcription failure, overlap is the lever. Start from the held-out overlap conditions and make them harsher; do not expect band-limiting, codecs or stationary noise to do the work. A detector trained only on the mild conditions never sees the failure it is meant to cover.

• 🤗 Model: NagaYu/bleep-0.09b
• 🎛️ Demo: NagaYu/bleep

What a row contains

utt_id, voice_key, condition, duration, subsets, and three parallel arrays — span_starts, span_ends, span_labels.

There is no transcript field, and no field containing what was said. The corpus annotates where sensitive speech occurs and what kind it is. That is a deliberate limit: a public dataset of spoken account numbers, even invented ones, is a worse artefact than one that only marks their positions.

Two configs, because of a licensing split

config	speech	audio included?	why
features	macOS system voices — realistic, 24 English speakers	❌ 64-band log-mel (fp16) + annotations	Apple's system voices are licensed for use, not for republication as a corpus. Features are what the model consumes anyway, and are not invertible to intelligible speech in any practical sense (magnitude only, no phase, 64 bands).
audio	Bleep's own numpy source-filter synthesiser	✅ full 16 kHz waveforms	We own the output outright, so it is freely redistributable. The speech is pseudo-speech: correct syllable rhythm, formant structure and prosody, but not intelligible words.

The published model was trained on features. The audio config is the same generator, conditions, splits and seed, rendered with the synthesiser we own — so you can listen to it, re-derive features, or use it to test a pipeline end to end.

from datasets import load_dataset

feats = load_dataset("NagaYu/bleep-spans", "features", split="test")
audio = load_dataset("NagaYu/bleep-spans", "audio", split="test")

Splits

Three speaker-disjoint partitions — no voice appears in more than one — plus a fourth evaluation set that reuses the test speakers with degradation conditions withheld from training, so a drop there is attributable to the condition rather than the voice.

calib exists because a threshold chosen on the test set makes "recall at a fixed false-alarm rate" a fitted parameter rather than a measurement — and choosing it on the training speakers would be optimistic in a subtler way, since the score distribution of a seen speaker is not that of an unseen one.

features config (13.15 hours)

split	utterances	hours	spans	voices	negatives
train	2200	9.15	2828	13	591
calib	250	0.94	278	4	37
test	550	1.99	641	7	77
test_heldout	300	1.07	345	7	39

audio config (14.64 hours)

split	utterances	hours	spans	voices	negatives
train	2200	10.16	2796	13	580
calib	250	1.02	288	4	32
test	550	2.26	629	7	67
test_heldout	300	1.21	345	7	38

Categories

• DIGIT_SEQUENCE
• GROUPED_DIGITS
• SPELLED_OUT
• DATETIME
• ADDRESS
• PERSON_NAME
• IDENTIFIER
• MONEY

Roughly balanced; on the test split: DIGIT_SEQUENCE 79, GROUPED_DIGITS 87, SPELLED_OUT 96, DATETIME 94, ADDRESS 63, PERSON_NAME 76, IDENTIFIER 85, MONEY 61.

The four weighted most heavily — GROUPED_DIGITS, SPELLED_OUT, DATETIME, IDENTIFIER — are the reported failure modes of transcript-based detection: digits read in chunks, names spelled letter by letter, relative time expressions, and identifiers spoken as separate elements. Utterances carry subsets tags (hard:grouped_digits, hard:spelled, hard:overlap, …) so these can be sliced directly.

How the labels are exact

Each utterance is rendered segment by segment — carrier speech, then the sensitive item, then more carrier — and concatenated. A span is the sample range the sensitive segment occupies by construction. There is no forced aligner in the loop and therefore no alignment error in the labels.

Degradation conditions

All strictly time-preserving, so the labels stay valid. Anything that stretched or delayed the signal would silently desynchronise every span, and nothing downstream would notice.

In training (8): clean, telephony_8k, noisy_babble_10db, noisy_street_5db, reverb_moderate, mic_variation, overlap_6db, telephony_noisy

Held out (5): packet_loss_voip, wind_noise_3db, reverb_heavy_phone, overlap_0db_three_way, telephony_overlap_3db (the last two were moved into training for the model's retrain; see below)

Two controls that protect a benchmark built on this corpus

Announcing-carrier decorrelation. Sensitive items are usually announced ("the account number is …"). If every announcement were followed by a sensitive item, a model could score well by detecting the announcement and never modelling the sensitive region — and a benchmark would report that as success. 35% of announcing carriers are followed by a harmless continuation instead.

Join-artefact decorrelation. Concatenating separately-rendered segments makes every segment boundary a potential artefact. If joins only ever occurred beside sensitive material, "a join is here" would itself predict sensitivity. Carrier phrases are therefore split into separately-rendered clauses too, so joins appear everywhere, including inside negative utterances. Measured ratio of non-sensitive to sensitive-adjacent joins: 1.01.

Known limitations

It is synthetic. Synthetic speech has narrower prosodic and disfluency variety than real conversation: no false starts, no laughter, no emotional range. Recall measured on it is an estimate of recall on real audio, not a measurement of it.

The training degradations are too mild. See the section above: on the training conditions the ASR baseline is flat (0.818 clean, 0.857 at 8 kHz), and it collapses only under overlap (0.271 at 0 dB three-way, 0.490 at narrowband-plus-overlap).

Concatenation removes coarticulation. Utterances are rendered segment-by-segment and concatenated, so there is no coarticulation across a carrier/sensitive boundary. Mitigations: randomised 30-160 ms inter-segment gaps, a 15 ms crossfade at each join, and the fact that every evaluation condition applies reverberation, noise and/or codec processing after concatenation, which smears join artefacts. A residual concern remains that the tagger could key on join artefacts rather than on speech; test_synth.py::test_join_artifacts_are_not_a_cue checks that carrier/carrier joins (present in every utterance, including negatives) are as frequent as carrier/sensitive joins, so a join is not by itself evidence of sensitivity.

English only. Japanese generators exist in the code and are wired end to end, but are not represented here and are not benchmarked.

What changed for the model's retrain (and what did not)

The files in this repository are unchanged. For the model's one convergence retrain, three splits were built in addition to this corpus, with scripts/build_corpus_v2.py in the GitHub repository. They are rendered with the macOS voices like the features config, and are not uploaded here:

split	utterances	speakers	conditions	purpose
train_harsh	1100 (550 per condition)	the 13 train voices	overlap_0db_three_way, telephony_overlap_3db	the two harsh held-out conditions, moved into training
dev_harsh	250	the 4 calib voices	the same two	early stopping sees the harsh conditions too
test_heldout_v2	300	the 7 test voices	music_bed_5db, lowbitrate_codec, hard_clipping, overlap_minus3db, agc_dropout	five conditions neither model version trained on

In consequence, for the retrained model two of test_heldout's five conditions are trained conditions (on unseen speakers), and test_heldout_v2 is the held-out set in the original sense. The eight training conditions and every split in this repository are as published.

Provenance

Generated by scripts/build_dataset.py (seed 20260918). Names, addresses and numbers are invented from word lists; any collision with a real person is coincidental and carries no information, because the strings exist only to drive a speech synthesiser and are discarded before any artefact is written.

A correction on reproducibility. An earlier version of this card said the build is deterministic. For the published train, calib and test rows that is not true: they were rendered under a per-process salted hash() seed and later re-stamped by a resumed build, so they cannot be regenerated exactly from the recorded seed. The published files are the reference. The builder now derives split seeds with zlib.crc32, and the retrain's additional splits were checked against their recorded seeds: for about 40 rows per split, the generation plan (template, speaking rate, voice) was re-derived from the seed and matched.

Licence

Apache-2.0 for the annotations, the audio config, and the generator code.