You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Humaneness Voice Large: intense bucket adapters

Model identity and current manuals

The specialist-study parent included in this repository is now also released as Humaneness Voice Large 1.6 Alpha: 1.5 plus final RealSpeech rank 4, already merged. Attach only one new bucket adapter; do not reload its ancestors.

Complete 1.6 + specialist context manual · Original 1.5 context manual · JSON routes · Interactive manual and audio. All 40 emotions, 15 talking styles and 14 edge cases have documented starting recipes; 72 matched reference-conditioned decisions preserve their controls and failed gates. Nine recommend a specialist, 45 use manual-only settings and 18 retain standard fallbacks. Existing license notices and access gates remain unchanged.

All 72 supported bucket adapters and their matched evaluations are complete. This repository uses a single already-merged Humaneness Voice Large 1.5 Alpha + final real-speech rank-4 parent. Its source rank4 checkpoint is step552. The fresh per-bucket adapters are rank32, with alpha32 and a native FP32 residual. Do not attach the old cumulative RL adapter or rank4 again. The included base already contains them.

These are supervised-fine-tuning adapters, not another GRPO run. Only individually audio-reviewed passing targets enter training; the larger raw-score candidate pool is not blindly used. Every supported language in a bucket is pooled and globally shuffled. Training uses three epochs, first-epoch warmup to2e-5 and cosine decay to2e-6, frozen parent weights and attention adapters in both semantic backbone and Talker. All optimizer/RNG/sampler/schedule checkpoints stay locally available. Fresh-process replay of an uninterrupted canary is a prerequisite.

The parent has a semantic backbone and local autoregressive Talker. It predicts12 MOSS-v2 codebook indices for each80ms frame,150 indices/s, decoded to48kHz mono. The LoRA rank is not the codec channel count. Native .a/.b adapter tensors are not PEFT files.

Listening and interpretation

Matched five-language listening grid · Gated dataset. The grid compares the same merged base with no new adapter against rank32 scales0.5,0.75,1.0. In the original standard-prompt section, these are residual scales, not CFG, and CFG is fixed at 1. The newer manual-comparison section separately sweeps CFG from 1 to 5 and compares matched no-adapter controls; use its current recommendations above. Prompt/text/seed/reference/frame budget are matched. All seeds are retained. Zero-support languages are labeled transfer evaluations, not training coverage. Scores are predicted diagnostics, not listener MOS.

S_MONO means monologue, not monotone. VOLT means emotional volatility, not loudness. Unsupported/insufficiently verified buckets have no fabricated checkpoint.

Use

Install the parent model dependencies, PyTorch, transformers, safetensors, huggingface_hub, numpy and soundfile. Access approval is required. Tested hardware is JUPITER GH200; FP32 frozen weights plus BF16 compute can require substantial GPU memory. The reference must be one clean speaker, another utterance,5.04–14.96 seconds. Never reinstate the historical2.96-second cap.

python infer_bucket.py --model laion/Humaneness-Voice-Large-1.5-Intense-Bucket-LoRAs   --bucket emo_Anger --scale 0.75 --language en   --text "The door is still open. Tell me when you are ready."   --prompt-file prompt.txt --reference-wav reference.wav --output prediction.wav

The example bucket must actually be in the completed inventory below. Use --scale 0 for the common merged baseline. The script applies the new adapter exactly once. prompt.txt contains English GENERAL/SCRIPT directions; --text contains only spoken words. Durations belong outside direction parentheses, e.g. (strong anger) [3.2 seconds duration] The door is still open. The base repository and public prompting manuals explain the trained wrapper. Training source and Slurm recipes are under training/jupiter/; machine-specific paths are explicit, not a claim of a universally portable one-command trainer.

No new collection license is assigned here; existing source/model notices remain applicable. Intended for research and authorized acting/voice work. Do not use for deceptive impersonation. Validate quality, identity and language behavior before deployment.

Completed adapters

emo_Affection
emo_Amusement
emo_Anger
emo_Astonishment_Surprise
emo_Awe
emo_Bitterness
emo_Confusion
emo_Contemplation
emo_Contempt
emo_Contentment
emo_Disappointment
emo_Disgust
emo_Distress
emo_Elation
emo_Fatigue_Exhaustion
emo_Fear
emo_Helplessness
emo_Hope_Enthusiasm_Optimism
emo_Impatience_and_Irritability
emo_Interest
emo_Longing
emo_Pain
emo_Relief
emo_Sadness
emo_Sourness
emo_Teasing
emo_Thankfulness_Gratitude
vn_S_ASMR_reg
vn_S_AUTH_reg
vn_S_CASU_reg
vn_S_CONV_reg
vn_S_DRAM_reg
vn_S_FORM_reg
vn_S_MONO_reg
vn_S_NARR_reg
vn_S_PLAY_reg
vn_S_STRY_reg
vn_VULN_reg_high
vn_VULN_reg_low
vn_STNC_reg_high
vn_STNC_reg_low
vn_AROU_reg_high
vn_AROU_reg_low
vn_VOLT_reg_high
vn_VOLT_reg_low
vn_VALN_reg_high
vn_VALN_reg_low
vn_WARM_reg_high
vn_TENS_reg_high
vn_TEMP_reg_high
casual_not_formal
formal_not_casual

Supplemental adapters for initially sparse buckets

These fresh rank32 adapters use the same included1.5Alpha+merged-rank4 parent. They are not stacked on any of the other bucket adapters. The original52 adapters are unchanged. Strongest-tail discovery was expanded for under-calibrated EARS heads; the independent per-audio intensity/quality gates and minimum training support were preserved. Each supported bucket pools its accepted languages for three epochs,first-epoch warmup to2e-5,then cosine decay. Code and operational recipes are in training/gap_repair/; the original training code remains in training/jupiter/.