laion/Humaneness-Voice-Large-1.5-Alpha

Humaneness Voice Large 1.5 Alpha

Expressive, multilingual text-to-speech with optional speaker-reference audio. This is the completed 20% Content Enjoyment / 40% emotion matching / 40% genuineness reinforcement-learning continuation. It is our preferred current listening checkpoint for a useful balance of naturalness and enjoyment.

Listen to the matched four-model comparison. The 20/40/40 column is exactly the 1.5 adapter released here. The page contains the same 70 generated cases per model, in English, German, French, Dutch and Spanish, with and without speaker references. Earlier columns are retained.

Everything needed for the model is included

Directory	What it actually contains
base_model/	Unmerged, unchanged Large 1.3 SFT weights, tokenizer and custom runtime
adapter/	Final cumulative 1.5 rank-128 LoRA, containing both RL phases
infer.py	Default loader: included base + final adapter, applied exactly once
training/	Portable continuation trainer, rewards, numerical objective and configuration
stats/	Aggregate evaluations, all 2,000 continuation optimizer steps and group curves

The base comes from Large 1.3 Alpha, pinned revision f4cee303eb362af4cf3d205e9fe0198288564398. Its original weights are duplicated here for convenience; base_model alone is not 1.5. The final adapter already includes the 1.4 RL phase. Do not stack the 1.4 adapter, and do not load this adapter on a base with 1.4 already merged. No rounded merged-weight substitute is used in this release. The external MOSS codec is downloaded separately and retains its upstream terms.

Quick start

The tested software stack is Python 3.13, PyTorch 2.9.1 with CUDA 12.8, Transformers 5.14.1 and Accelerate 1.14.0, on GH200 GPUs. Install a CUDA PyTorch build appropriate for your server first. The reviewed custom MOSS runtime requires trust_remote_code=True; do not use an unreviewed changed revision.

hf download laion/Humaneness-Voice-Large-1.5-Alpha --local-dir humane-large15
cd humane-large15
pip install -r requirements.txt
python infer.py --language en --seconds 7 --seed 777 \
  --text 'I thought you had forgotten. What a wonderful surprise.' \
  --prompt 'GENERAL: warm, spontaneous, quietly delighted.\nSCRIPT:\n(warm relief) [3.0 seconds duration] I thought you had forgotten.\n(brighter gratitude) [4.0 seconds duration] What a wonderful surprise.' \
  --output example.wav

For speaker conditioning add --reference-wav clean-reference.wav. Use clean, single-speaker speech 5.04–14.96 seconds long. Longer recordings are cropped to 14.96 seconds before encoding; there is no accidental three-second cap. Use a different utterance from the requested spoken text. Do not include music, multiple speakers or large silent regions. Good references help, but exact identity preservation is not guaranteed.

import infer
model, processor, base_api = infer.load(device='cuda:0')
base_api.load_codec(processor, 'OpenMOSS-Team/MOSS-Audio-Tokenizer-v2')
message = base_api.build_user(processor,
    text='I am so glad you came back.',
    prompt='GENERAL: relieved, affectionate, natural.\nSCRIPT:\n(quiet relief) [4.0 seconds duration] I am so glad you came back.',
    frames=50, language='en')
codes = base_api.generate(model, processor, message, seed=777, max_frames=240)
waveform, sample_rate = base_api.decode(processor, codes)

See PROMPTING.md for complete English/German examples, sentence-level durations, pauses, vocal bursts, reference versus voice-name conditioning, and common formatting mistakes. Spoken text and performance directions are separate fields: do not accidentally put director instructions inside the literal text that should be spoken.

Adapter format matters

This is a custom FP32 a/b attention adapter, not PEFT format. Use the supplied loader rather than PeftModel.from_pretrained. Rank=alpha=128; residual scale=1. The 292 tensors contain 96,337,920 trainable scalars. They adapt semantic q/k/v/o projections and local-Talker attention c_attn/c_proj. The base executes in BF16 for deployment, while the residual multiplication stays FP32. The public listening page uses this unmerged path.

Final adapter SHA-256: 7f376cf9b9605f8ae8d27a8ebe521f17459d91bbfabfcf3d883f9381b5d67549. All published files have checksums and sizes in MANIFEST.json.

What the architecture does

The MOSS Local Transformer family model has 4,129,990,144 unique base parameters: 4,022,468,096 in the semantic transformer and 107,522,048 in the local Talker. The Qwen3-family semantic transformer has 36 layers, hidden size 2,560, 32 attention heads and 8 key/value heads. The one-layer local Talker also uses hidden size 2,560.

For each 80 ms frame, the semantic transformer provides context and the local Talker predicts 12 codec indices, each from a vocabulary of 1,024. That is 12.5 frames/second × 12 = 150 indices/second, not 12 indices/second. The external MOSS Audio Tokenizer v2 decodes them to 48-kHz mono audio. The current checkpoint uses the first twelve codec codebooks; it does not generate all 32 codebooks of the full codec.

Matched results — all 70 cases retained

Model	Content Enjoyment ↑	Genuineness ↑	Mean per-clip WER ↓	Approx. emotion raw MSE ↓
SFT 1.2	6.235	4.244	1.133%	1.838
SFT 1.3, included base	6.259	4.200	3.238%	1.833
RL 1.4, 33/33/34	6.328	4.299	2.743%	1.793
Humaneness Voice Large 1.5 Alpha	6.331	4.370	1.942%	1.791

Compared with 1.4, average genuineness rises and WER falls; CE and the approximate emotion diagnostic barely change. SFT 1.2 still has the lowest WER. These are learned proxy scores, not human MOS or proof of statistical significance. CE and genuineness come from Humaneness EARS Medium; ASR is Parakeet v3. WER is the arithmetic mean of clip-level error rates, not corpus-token-weighted WER.

The emotion column uses fixed prompt-derived scene-level targets, with two named emotion heads per scene. They are approximate intended intensities, not a ground-truth performance and not the source-audio targets used during RL training. There is no speaker-similarity claim for this comparison. All original prompts, seeds, reference crops and frame budgets are matched; no best-of-N, new seeds, clipping or rejected-output exclusion improves the table. Full aggregates and sample standard deviations include the reference/no-reference breakdown and the complete diagnostic definition.

Training history and reproduction

The frozen SFT 1.3 base first received 1,452 successful rank-128 RL updates with rewards 33/33/34. This release continues that same cumulative adapter for 2,000 additional successful updates with 20/40/40, for 3,452 total. The continuation uses paired groups of 16 candidates with and without reference audio, across 32 GPUs. It attempts 2,088 groups, rejects 88 under unchanged quality gates, and saves 42 full checkpoints. A new 5%-warmup/cosine schedule uses peak LR 3e-6 and floor 3e-7. Full AdamW moments were explicitly migrated from the first phase; this phase change is not an exact old-phase resume. An infrastructure interruption after update 520 was later resumed exactly within the unchanged 32-rank phase, restoring optimizer and all rank RNG states.

The experimental objective combines a VESPO sequence-importance kernel with GDPO-style separate reward-component normalization. It is not a stock TRL GRPO trainer, and it is not the flow-matching GROW algorithm. The reward matches the two strongest fixed emotions of the source target recording, never whichever emotions happen to score highest in a generated candidate or speaker reference.

TRAINING.md explains rewards, prepared-input schema, exact hyperparameters, distributed launch, warm-start versus exact resume, and what can and cannot be reproduced from this public package. New experiments can warm-start from the final adapter with a fresh optimizer. The original private prepared input map, raw recordings and optimizer pickle are not distributed. Training inputs are described accurately as balanced real/synthetic speech across the five listed languages; provider-specific labels and identifiers are omitted. Method reproduction with compatible inputs is supported, but byte-identical original-data reproduction is not claimed.

License and responsible use

New weights and release materials are CC BY 4.0; see LICENSE and NOTICE.md. Included base weights retain their original CC BY 4.0 declaration and lineage notices. Original software headers and separate dependency terms remain applicable; CC BY does not relicense third-party code or grant rights over someone else's identity.

This alpha is primarily for research and responsibly licensed dataset augmentation. Test pronunciation, identity consistency, bias, timing and audio quality before downstream deployment. Obtain consent and necessary rights for voice conditioning; do not use the model for deceptive impersonation, fraud or misleading attributed speech. LAION contributors do not endorse generated statements or assume responsibility for downstream use. These are responsible- use recommendations, not additional legal restrictions replacing CC BY 4.0.