laion/Humaneness-Voice-Large-1.6-Alpha

Humaneness Voice Large 1.6 Alpha

Expressive multilingual text-to-speech with sentence-level acting directions and optional speaker-reference audio. 1.6 is Humaneness Voice Large 1.5 Alpha plus the final RealSpeech rank-4 adapter, merged once at strength 1.0. The complete merged model is included. There is no additional training for this version number and no specialist adapter enabled by default.

Manuals, demos and adapter navigation

What you need	Start here
Complete context manual for this model	1.6 Alpha: all recipes and specialist recommendations
Agent/tool integration	Agent entry · compact JSON registry · recipe compiler
Original policy, before RealSpeech rank 4	1.5 Original manual · 1.5 model
Detailed recipes and audio evidence	Prompting manual Space
Matched specialist tests on this merged parent	Standard / manual / CFG / adapter listening comparisons
Earlier rank comparison	RealSpeech listening study
Additional adapters	Specialist rank-32 family · historical RealSpeech ranks

The self-contained manuals cover 40 emotions, 15 talking styles and 14 edge cases per policy. The 1.6 manual adds 72 reference-conditioned decisions: 9 specialist adapters, 45 manual-only settings and 18 standard fallbacks. Failed gates and unvalidated cases remain explicit. A recipe for every target is not a claim that every target can be produced reliably.

Quick start

The tested stack is Python 3.13, PyTorch 2.9.1/CUDA 12.8, Transformers 5.14.1 and Accelerate 1.14.0 on GH200. Install a CUDA PyTorch build for your hardware first. Review and pin this repository's custom model code before using trust_remote_code=True.

hf download laion/Humaneness-Voice-Large-1.6-Alpha --local-dir humaneness16
cd humaneness16
pip install -r requirements.txt
python infer.py --language en --seconds 10.2 --seed 777   --text 'I thought you had forgotten. You remembered every little detail. I am so glad you came back.'   --prompt-file examples/prompt.txt --output example.wav

Add --reference-wav clean-reference.wav for speaker conditioning. Use clean single-speaker speech from another utterance, at least 5.04 seconds long. The waveform is cropped to 14.96 seconds before encoding, with 63–187 codec frames retained. No old three-second cap is reinstated. Alternatively use --voice-name with a registered alias from the voice catalog; an alias and a reference are mutually exclusive.

Release verification: all four weight shards match the previously validated merged parent; all 138 conditional/neutral recipe pairs and 72 effective route prompts were reconstructed exactly. The additional public-loader GPU canaries passed for no-reference CFG 1 and long-reference CFG 2.5 generation; results are recorded in validation.json. This is separate from the completed scientific evaluations.

The root loader keeps merged weights in FP32 and computes in BF16, matching the specialist study. Weight storage is approximately 16.5 GB; generation additionally needs model activations, KV cache and the external codec. The matched two-seed study used approximately 32 GB peak GPU memory, but this is not a maximum for every sequence length. Codec weights are downloaded separately from MOSS Audio Tokenizer v2.

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

Prompt format

Keep the spoken words in --text. Performance directions are English, even for German, French, Dutch or Spanish speech. The same words appear in SCRIPT, in the same order:

GENERAL: one adult speaker, intimate, spontaneous, warmly amused.
SCRIPT:
(soft amusement) [3.2 seconds duration] I thought you had forgotten.
[0.3 seconds pause]
(chuckle, 0.4 seconds)
(brighter, affectionate) [3.8 seconds duration] You remembered every little detail.
(quiet gratitude) [3.2 seconds duration] I am so glad you came back.

Round brackets direct a sentence or request a nonverbal event. Square brackets hold speech-duration or pause requests. Speech duration is outside the round-bracket direction. Add pauses/events to the total --seconds budget. Do not insert stage directions into the literal spoken text. An explicit duration is a request, not a guarantee of exact timing. See PROMPTING.md and the complete manual for P08 GENERAL-only and P09 inline-only recipes.

Choosing CFG and optional specialist LoRAs

CFG and adapter strength are different controls. CFG combines conditional and neutral logits, with separate caches following the same sampled audio tokens: neutral + CFG * (conditional - neutral). The conditional stop decision is unchanged. Higher CFG is not universally better; use the recorded per-target setting.

The repository's infer_recipe.py selects the actual recorded prompt, CFG and, when justified, specialist strength. For example, the reference-conditioned anger route uses rank 32 at strength 0.75 and CFG 2.5:

python infer_recipe.py --target emo_Anger --language en   --sentences-json examples/sentences.json --reference-wav clean-reference.wav   --seed 1701 --output anger.wav

Use scene-appropriate sentences in your own JSON file; the supplied three sentences only demonstrate the input structure. Add --compile-only to inspect instructions without inference. Pure model recipes use --recipe, with IDs in the compact JSON. Diagnostic recipes require explicit --allow-diagnostic.

Specialist weights live in the manually gated Intense-Bucket-LoRAs repository. Request access there and authenticate locally when using a specialist. The helper pins each adapter revision and verifies its checksum. Manual-only and standard-fallback routes need no gated weights. These adapters are native a/b tensors, not PEFT format. Their parent is exactly the merged model released here, despite the historical “1.5” in that repository's name.

Do not attach the old rank-128 RL adapter or the RealSpeech rank-4 adapter again. Do not attach these specialists to unmodified 1.5. Do not stack multiple specialists without a new evaluation. Specialist gains were tested with references in five languages, not without references or on every possible speaker.

Model architecture and lineage

The base has 4,129,990,144 unique parameters: 4,022,468,096 in the Qwen3-family semantic transformer and 107,522,048 in its local Talker. The semantic model has 36 layers, width 2,560, 32 attention heads and 8 KV heads; the one-layer Talker has width 2,560. Every 80 ms, the Talker predicts twelve codec indices from vocabularies of 1,024. This is 150 indices per second, decoded to 48-kHz mono; it uses twelve of the codec's available codebooks.

1. The included 1.5 lineage starts from Large 1.3 SFT and its cumulative rank-128 RL adapter. The final RL phase rewards 20% Content Enjoyment, 40% source-top-two emotion matching and 40% genuineness.
2. A rank-4 adapter was trained for three epochs on approximately 25 hours of selected real speech: five hours per language, 5,860 unique clips. Its final checkpoint is update 552.
3. Both ancestors were merged once into FP32 base weights. These already existed as the parent used in the specialist comparisons. This release gives that parent the name 1.6 Alpha.

The later rank-32 RL continuation is a separate experiment and is not included. The merge record gives the source adapter checksum, four final weight hashes and algebraic validation (maximum tested linear error 3.34e-6).

What the evidence does and does not show

The matched specialist study uses this exact merged parent, unseen text and two seeds in five languages, with one fixed reference voice per language: 50 takes per condition and target. It compares standard prompts, manual prompts, matched-CFG controls and new bucket adapters. The manual retains all quality gates, failures and control results. The nine recommended specialists show a qualified target improvement in this study; they are not universal guarantees.

Earlier model-specific rank-4 prompting and 70-take real-speech comparisons used a separately loaded adapter on a BF16 base. Merging changes rounding, so those audio files remain labeled as historical unmerged rank-4 evidence, not freshly generated 1.6 samples. Their recipes are inherited starting points. Earlier rank-4 WER was 0.971% versus 1.942% for original 1.5, while CE barely changed and original 1.5 retained higher mean genuineness. Those development measurements do not establish that 1.6 is universally more natural. Listen to the linked matched comparisons for your use case.

Training and reproducibility

See TRAINING.md for corpus selection, schedule, loss, reference conditioning, code and reproduction boundaries. Rank-4 training statistics preserve the learning curve and validation results; selection statistics document the corpus. The portable trainer starts a new native adapter on the selected parent and saves optimizer/RNG states. It is not a bit-exact recreation of the original distributed run.

Path	Contents
base_model/	Full merged 1.6 weights, model code, tokenizer, aliases, upstream notices
infer.py, tools/, native_cfg.py	Model loading, reference handling and true logit CFG
infer_recipe.py, native_adapters.py	Deterministic recipe selection and optional specialist loading
manuals/, AGENT_ENTRY.md	Self-contained human/agent manuals, JSON registries and token counts
training/, TRAINING.md	Portable SFT code and historical reproduction instructions
stats/, MANIFEST.json	Selection/training/merge evidence and file hashes

License and responsible use

New model contributions, code and documentation are released under CC BY 4.0. Preserve the upstream notices, attribution and applicable dependency licenses. This release does not change the manually gated specialist collection's terms or grant rights over third-party training/reference audio.

Intended uses include research, authorized voice acting and dataset augmentation. Test language, speaker fidelity, safety and failure modes before deployment. Use reference voices only with appropriate permission; do not use the model for deceptive impersonation. These are usage recommendations, not extra restrictions added to the CC BY license. LAION does not endorse misuse or guarantee any particular output quality.