MOSS Explicitness (NSFW) LoRAs — dose-sweep grid

Small trainable add-ons (LoRA adapters) for the expressive text-to-speech model laion/moss-tts-local-transformer-4.55b-voice-acting-v2 that make a synthetic voice sound more sexually suggestive when it speaks. Each row below is one such adapter; the players let you hear the effect and the numbers say how strong it is and what it costs in quality.

🔞 18+ · Research only. Everything here is fully synthetic: the training clips and all samples were generated by AI text-to-speech systems (a drama-dialogue synthesis pipeline and a commercial TTS model) — there are no recordings of real people. Scores are produced automatically by machine-learning models, not human raters. The content is suggestive but intended to stay within what ordinary adults would find acceptable.

▸ What did we do, and what do all these abbreviations mean? (plain-English explainer)

The idea. The base voice model can already act out emotions. We wanted an add-on that pushes the delivery toward sexually suggestive without retraining the whole model. A LoRA is a tiny "style patch" you blend into the model at a chosen strength (the dose, e.g. 0.5× … 1.25×). More dose = stronger effect, but usually more side-effects.

How we picked the training clips. We scored a large pool of synthetic voice clips and kept, in two steps, the ones that were both aroused-sounding and good-sounding: (1) keep the top 10% by Sexual Lust; (2) of those, keep the half that also had high explicitness and a natural vocal quality (for the "aesthetic" variant we also weighted pleasantness). The variants differ only in the data split used:

A · aesthetic mix — the aesthetics-filtered pool (nicest-sounding). The recommended default.
B · adult — the same recipe but only the adult-voice split.
C · A1-suggestive — only the suggestive (non-explicit) drama split.
D · raw lust — the top 5% by Sexual Lust with no quality filter (the raw contrast), trained at three sizes (rank 16/32/64) to find the best.

The scores (what the columns mean):

Sexual_Lust — the key metric: how sexually aroused / turned-on the voice sounds (from the EmoNet emotion model). Higher = more. This is what the LoRAs are meant to raise.
BLEND (vocal-burst blend) — how naturally little non-speech sounds (moans, gasps, sighs) are woven into the speech. Higher = more natural.
ESTH (aesthetics) — how pleasant / high-quality the voice is to listen to.
GENU (genuineness) — how genuine / non-fake the emotion sounds.
WER (word-error-rate) — how unintelligible the words are (0 = every word clear, 1 = unintelligible). Lower is better; suggestive, breathy delivery naturally raises it a bit.
EXPL — a generic "content-appropriateness" score. It turned out to be a poor guide here (it is inverted — low = explicit — and barely tracks Sexual_Lust, correlation only −0.15), which is exactly why we switched to the Sexual_Lust-based recipe above.

Every number is a cohort mean over 8 generated samples (not one lucky take). "Baseline" = the model with no LoRA. Δ = change vs. baseline. Note: Google Gemini refused to rate this audio (safety filter), so perceptual validation relies on the Sexual_Lust score plus listening.

Baseline (no LoRA): Sexual_Lust -0.11 · BLEND 0.42 · ESTH 2.80 · GENU 0.18 · WER 0.87

Best dose per adapter

adapter	best dose	Sexual_Lust	BLEND	ESTH	GENU	WER
A · Aesthetic mix (r32)	1.25×	+0.53 (Δ+0.65)	0.43	2.96	0.17	0.61
B · Adult (r32)	1.0×	+0.47 (Δ+0.58)	0.42	2.88	0.18	0.52
C · A1-suggestive (r32)	1.25×	+0.43 (Δ+0.54)	0.32	2.93	0.18	0.87
D · Raw lust (r16)	1.0×	+0.53 (Δ+0.64)	0.45	2.95	0.20	0.62
D · Raw lust (r32)	1.25×	+0.63 (Δ+0.74)	0.42	2.74	0.18	0.49
D · Raw lust (r64)	1.25×	+0.60 (Δ+0.71)	0.39	2.89	0.20	0.64

Take-aways: rank-32 wins the D rank-sweep (Sexual_Lust +0.74 @1.25×); A/aesthetic keeps BLEND+ESTH highest at high lust (NSFW and pleasant); higher dose = more lust, at a cost to genuineness & WER (the suggestive carrier text is inherently less crisply articulated).

A · Aesthetic mix (r32)

Sexual_Lust top-10% → re-ranked by EXPL+BLEND+ESTH. NSFW and nice-sounding — the recommended default.

best dose @1.25 → Sexual_Lust +0.53 (Δ+0.65)

dose	0.5×	0.75×	1.0×	1.25×
Sexual_Lust	+0.31 Δ+0.42 BL 0.50·ES 2.52·GE 0.21·WER 0.77	+0.22 Δ+0.33 BL 0.52·ES 2.50·GE 0.21·WER 0.52	+0.29 Δ+0.40 BL 0.45·ES 2.98·GE 0.19·WER 0.73	+0.53 Δ+0.65 BL 0.43·ES 2.96·GE 0.17·WER 0.61

listen @1.0×:

neutral_s40

neutral_s41

B · Adult (r32)

adult dataset only, same aesthetic re-rank.

best dose @1.0 → Sexual_Lust +0.47 (Δ+0.58)

dose	0.5×	0.75×	1.0×	1.25×
Sexual_Lust	+0.33 Δ+0.44 BL 0.49·ES 2.77·GE 0.20·WER 0.76	+0.04 Δ+0.15 BL 0.43·ES 2.85·GE 0.19·WER 0.64	+0.47 Δ+0.58 BL 0.42·ES 2.88·GE 0.18·WER 0.52	+0.40 Δ+0.51 BL 0.43·ES 2.78·GE 0.18·WER 0.98

listen @1.0×:

neutral_s40

neutral_s41

C · A1-suggestive (r32)

additional-data-a1-plus only — suggestive, non-adult source.

best dose @1.25 → Sexual_Lust +0.43 (Δ+0.54)

dose	0.5×	0.75×	1.0×	1.25×
Sexual_Lust	-0.08 Δ+0.03 BL 0.44·ES 2.87·GE 0.17·WER 0.85	+0.20 Δ+0.31 BL 0.46·ES 2.86·GE 0.17·WER 1.00	+0.38 Δ+0.49 BL 0.40·ES 3.04·GE 0.17·WER 0.87	+0.43 Δ+0.54 BL 0.32·ES 2.93·GE 0.18·WER 0.87

listen @1.0×:

neutral_s40

neutral_s41

D · Raw lust (r16)

top-5% pure Sexual_Lust, NO aesthetic filter — the raw contrast.

best dose @1.0 → Sexual_Lust +0.53 (Δ+0.64)

dose	0.5×	0.75×	1.0×	1.25×
Sexual_Lust	+0.22 Δ+0.33 BL 0.46·ES 2.50·GE 0.21·WER 0.76	+0.08 Δ+0.19 BL 0.44·ES 2.79·GE 0.21·WER 0.76	+0.53 Δ+0.64 BL 0.45·ES 2.95·GE 0.20·WER 0.62	+0.35 Δ+0.47 BL 0.43·ES 2.96·GE 0.17·WER 0.87

listen @1.0×:

neutral_s40

neutral_s41

D · Raw lust (r32)

raw top-5% — best of the rank sweep (highest Sexual_Lust lift).

best dose @1.25 → Sexual_Lust +0.63 (Δ+0.74)

dose	0.5×	0.75×	1.0×	1.25×
Sexual_Lust	-0.02 Δ+0.09 BL 0.48·ES 2.36·GE 0.22·WER 0.75	+0.38 Δ+0.49 BL 0.42·ES 2.61·GE 0.22·WER 0.53	+0.16 Δ+0.27 BL 0.48·ES 2.94·GE 0.17·WER 0.90	+0.63 Δ+0.74 BL 0.42·ES 2.74·GE 0.18·WER 0.49

listen @1.0×:

neutral_s40

neutral_s41

D · Raw lust (r64)

raw top-5%, rank 64.

best dose @1.25 → Sexual_Lust +0.60 (Δ+0.71)

dose	0.5×	0.75×	1.0×	1.25×
Sexual_Lust	-0.00 Δ+0.11 BL 0.40·ES 2.71·GE 0.20·WER 0.74	+0.27 Δ+0.39 BL 0.42·ES 2.73·GE 0.17·WER 0.39	+0.22 Δ+0.33 BL 0.36·ES 2.52·GE 0.22·WER 0.62	+0.60 Δ+0.71 BL 0.39·ES 2.89·GE 0.20·WER 0.64

listen @1.0×:

neutral_s40

neutral_s41

Adapters: TTS-AGI/moss-explicitness-loras (private). Data: TTS-AGI gemini-adult-voices + additional-data-a1-plus. Scoring: EmoNet Sexual_Lust · VoiceNet EXPL/ESTH · VoiceCLAP genu/blend · faster-whisper WER. Part of the voice-acting manual.