MOSS VoiceNet Dimension LoRAs

100 rank-32 LoRA adapters for laion/moss-tts-local-transformer-4.55b-voice-acting-v2. Each adapter pushes one of the 57 VoiceNet voice-performance dimensions in the high or low direction. Merge one into the base voice-acting checkpoint to steer that single dimension of the generated voice (e.g. warmer timbre, deeper chest resonance, faster tempo, more/less dramatic delivery).

This release: 57 __high + 43 __low = 100 adapters. 14 low-direction dimensions are still training and will be added: AGEV, ATCK, CHNK, EMPH, FOCS, R_MASK, SMTH, STRU, S_AUTH, S_CONV, S_NEWS, S_RANT, S_STRY, TEMP.

The VoiceNet dimensions

The 57 dimensions come from the VoiceNet-Ext taxonomy (Schumann et al., 2025 — arXiv:2505.20033), covering rhythm & timing, social dynamics, speaker identity, emotion & affect, physical production, spectral/timbral quality, temporal dynamics, recording conditions, resonance placement, and speaking style. Each dimension is scored on 7 ordinal levels (0–6).

code	dimension	high	low
`AGEV`	Voice Age	✅	⏳
`AROU`	Arousal	✅	✅
`ARSH`	Arousal Shift	✅	✅
`ATCK`	Attack	✅	⏳
`BKGN`	Background Noise	✅	✅
`BRGT`	Brightness	✅	✅
`CHNK`	Chunking	✅	⏳
`CLRT`	Articulation Clarity	✅	✅
`COGL`	Cognitive Load	✅	✅
`DARC`	Dynamic Arc	✅	✅
`DFLU`	Disfluency	✅	✅
`EMPH`	Emphasis	✅	⏳
`ESTH`	Esthetics	✅	✅
`EXPL`	Content Appropriateness (3-point Scale)	✅	✅
`FOCS`	Focus	✅	⏳
`FULL`	Fullness	✅	✅
`GEND`	Perceived Gender	✅	✅
`HARM`	Harmonicity	✅	✅
`METL`	Metallic Character	✅	✅
`RANG`	Pitch Range	✅	✅
`RCQL`	Recording Quality	✅	✅
`REGS`	Register	✅	✅
`RESP`	Respiration	✅	✅
`ROUG`	Roughness	✅	✅
`R_CHST`	Chest Resonance	✅	✅
`R_HEAD`	Head Resonance	✅	✅
`R_MASK`	Mask Resonance	✅	⏳
`R_MIXD`	Mixed Resonance	✅	✅
`R_NASL`	Nasal Resonance	✅	✅
`R_ORAL`	Oral Resonance	✅	✅
`R_THRT`	Throat Resonance	✅	✅
`SMTH`	Smoothness	✅	⏳
`STNC`	Stance	✅	✅
`STRU`	Structure	✅	⏳
`S_ASMR`	ASMR Style	✅	✅
`S_AUTH`	Authoritative Style	✅	⏳
`S_CART`	Cartoonish Style	✅	✅
`S_CASU`	Casual Style	✅	✅
`S_CONV`	Conversational Style	✅	⏳
`S_DRAM`	Dramatic Style	✅	✅
`S_FORM`	Formal Style	✅	✅
`S_MONO`	Monologue Style	✅	✅
`S_NARR`	Narrator Style	✅	✅
`S_NEWS`	Newsreader Style	✅	⏳
`S_PLAY`	Playful Style	✅	✅
`S_RANT`	Ranting/Angry Style	✅	⏳
`S_STRY`	Storytelling Style	✅	⏳
`S_TECH`	Teacher/Didactic Style	✅	✅
`S_WHIS`	Whisper-Talk Style	✅	✅
`TEMP`	Tempo	✅	⏳
`TENS`	Tension	✅	✅
`VALN`	Valence	✅	✅
`VALS`	Valence Shift	✅	✅
`VFLX`	Velocity Flux	✅	✅
`VOLT`	Volatility	✅	✅
`VULN`	Vulnerability	✅	✅
`WARM`	Warmth	✅	✅

Naming

vn_<CODE>__<high|low> — e.g. vn_WARM__high (warmer timbre), vn_R_CHST__high (more chest resonance), vn_TEMP__low (slower tempo).

How to use — merge into the v2 voice-acting checkpoint

These LoRAs are not standalone: they must be merged into the base model laion/moss-tts-local-transformer-4.55b-voice-acting-v2.

import torch
from transformers import AutoModel
from peft import PeftModel

base = AutoModel.from_pretrained(
    "laion/moss-tts-local-transformer-4.55b-voice-acting-v2",
    trust_remote_code=True, torch_dtype=torch.bfloat16).to("cuda").eval()

# load one dimension LoRA (subfolder = the adapter name)
pm = PeftModel.from_pretrained(base, "laion/moss-voicenet-dimension-loras",
                               subfolder="vn_WARM__high", adapter_name="L").eval()

# choose the MERGE STRENGTH (dose). 50% already delivers most of the effect; 100% is stronger; higher over-drives.
STRENGTH = 0.5
for m in pm.modules():
    sc = getattr(m, "scaling", None)
    if isinstance(sc, dict) and "L" in sc:
        sc["L"] = sc["L"] * STRENGTH
pm.base_model.set_adapter(["L"])

# ...then generate with the MOSS processor exactly as with the base voice-acting model...

Dose finding (from our sweep): averaged over all dimensions, 50% merge already reaches ~most of the full-strength shift; 100% pushes harder; beyond that the voice starts to over-drive. Tune per dimension.

Audio demos

Full showcase — every LoRA at 0 / 50 / 100 % (with the best-found steering prompt per dimension): https://laion-ai.github.io/moss-voicenet-lora-evolution/
High-vs-low contrast grid (same text, high vs low LoRA side by side): https://laion-ai.github.io/moss-voicenet-highlow-grid/

Provenance

Trained on VoiceNet-scored speech buckets with ≤25% game-character data (to reduce caricature bias), rank 32, on the voice-acting v2 checkpoint. VoiceNet scoring uses the LAION VoiceNet dimension predictors.

Downloads last month: -

Model tree for laion/moss-voicenet-dimension-loras

Base model

laion/moss-tts-local-transformer-4.55b-voice-acting-v2

Adapter

(4)

this model

Paper for laion/moss-voicenet-dimension-loras

EmoNet-Face: An Expert-Annotated Benchmark for Synthetic Emotion Recognition

Paper • 2505.20033 • Published May 26, 2025 • 4