Humaneness Voice Small 1.1 Alpha — RL Rank-128 v1
This experimental rank-128 attention LoRA adds group-relative reinforcement learning to the final Humaneness Voice Small 1.1 Alpha. This repository now includes the complete, exact frozen SFT5 base in base/ plus the separate rank-128 adapter. Download the whole repository and run its top-level infer.py: it loads both automatically. The files are not merged, and the adapter is not PEFT format. The upstream base-model repository and its default remain unchanged.
Listen and compare the actual final adapter. The corresponding Small/Large columns are labeled “RL-v1 (earlier recipe)” with the exact update count; nearby FP32/temperature-0.7 controls isolate the adapter from inference-setting changes. Complete audio and scores are public. This v1 used CE, VoiceCLAP and speaker identity rewards. It does not use the later proposed direct genuineness/top-emotion/top-talking-style reward.
Prompting: start here
Read the complete Prompting Guide for copyable GENERAL: / SCRIPT: examples, sentence durations, pauses, vocal bursts and reference-audio guidance. Pass the exact spoken words separately in --text; use --prompt for the matching annotated script and English performance instructions. The adapter loader accepts a reference, a known alias or neither; alias use was evaluation-only in this RL run. Use the top-level infer.py with its pinned bundled base.
Basic format:
GENERAL: Warm, relieved and naturally conversational.
SCRIPT:
(softly, with relief) [2.4 seconds duration] We finally made it home.
[0.4 seconds pause]
(warmly) [2.0 seconds duration] I am glad we are together.
Pass this structured block as --prompt and the matching spoken words, without direction or timing markup, as --text: We finally made it home. I am glad we are together. Use (chuckle, 0.5 seconds) for a vocal burst at the desired position. See PROMPTING_GUIDE.md for reference-audio and alias conditioning, frame budgets, additional examples and model-specific inference details.
Complete base + adapter bundle
| Item | Exact released component |
|---|---|
| Base | laion/Humaneness-Voice-Small-1.1-Alpha, SFT5 step 6,109 |
| Base files in this repository | base/checkpoints/step-00006109/model.safetensors, matching architecture/tokenizer assets, and the original base loader |
| Adapter | adapter.safetensors, rank128, 2,434 optimizer updates |
| Default inference | Top-level infer.py loads the included base and adapter, using the already-tested FP32/temperature0.7 path |
| Upstream codec | MOSS Audio Tokenizer v2, downloaded separately; reward judges are not needed for inference |
| Proofs | stats/bundled_base_verification.json, original reference/no-reference GPU canaries and matched207-case results |
See BUNDLE.md for layout, minimal commands, numerical safeguards and testing scope. Packaging does not change the adapter or its training: this is RL-v1, not the newer RL-v2 or SFT6 runs. Large stored base weights are about16.52GB and Small about2.86GB; use a roomy download/cache directory rather than a quota-limited home folder.
What changed, in simple terms
For each training prompt the policy produced eight different candidate recordings. Frozen scoring models assessed how enjoyable they sounded, whether delivery matched the performance directions, and—when a reference was provided—whether the speaker matched it. The training update makes relatively better candidates within that group more likely, while anchoring the policy to the starting model and known target speech. Only the adapter matrices change; the Alpha base stays frozen.
An 80-ms semantic frame conditions a local Talker, which predicts 12 MOSS-v2 residual codec indices per frame. That is 12.5 frames/s × 12 codebooks = 150 codec indices/s. The RL objective includes the hierarchical semantic/codec decisions; this is not an ordinary text-only TRL training call and not the flow-matching GROW algorithm used for Humaneness Voice Converter.
Download and inference
Install the tested stack from requirements.txt in a CUDA environment. Download this complete inference bundle:
hf download laion/Humaneness-Voice-Small-1.1-Alpha-RL-r128-v1 --local-dir adapter
python adapter/infer.py --adapter-dir adapter --language en \
--text 'We finally made it home.' \
--prompt 'GENERAL: A warm, relieved voice; relaxed and naturally conversational.
SCRIPT:
(relieved, softly exhaling, 3.2 seconds) We finally made it home.' \
--frames 75 --seed 777 --output example.wav
The frozen base files at revision b92b1f86e17344b33bf12f255767f589b785acee are already included in base/; there is no second TTS base-model download. infer.py prefers that local bundle and checks the base and adapter SHA-256 digests. The MOSS-v2 codec remains an upstream dependency and is obtained separately. For offline synthesis, also download OpenMOSS-Team/MOSS-Audio-Tokenizer-v2, then add --codec-path /path/to/codec. A deliberately separate pinned base can still be selected with --base-dir /path/to/base. An incomplete bundled download fails with a clear error rather than silently fetching another model.
For voice cloning, add --reference-wav clean-speaker.wav to the same command. Use a distinct, clean 5–15-second recording, preferably containing one speaker without music. The base CLI caps the waveform at 15 seconds before codec encoding; it does not use the historical 2.96-second limit. A shorter reference warns. No-reference mode has no hidden audio/name. --alias Alice is retained for compatibility with the base, but named-voice prompting was not trained by this RL run and is an evaluation-only condition.
These exact adapter tensor exports and loader passed GPU reproduction tests: both reference and no-reference generations reproduced the already published demo codec tokens exactly. See stats/public_loader_canary_*.json. Tested hardware: NVIDIA GH200; PyTorch 2.9.1+CUDA 12.8; Transformers 5.14.1. The released inference follows FP32 policy computation on BF16-rounded base weights, TF32 disabled, audio temperature 0.7, top-p 0.95, top-k 50. Arbitrarily casting the model to BF16 or using PEFT changes this contract and has not been validated here. Judge models are not required for synthesis.
RL reward and objective
| Training condition | Content Enjoyment | VoiceCLAP performance-text cosine | Orange speaker identity cosine |
|---|---|---|---|
| With reference | 40% | 30% | 30% |
| Without reference | 60% | 40% | Not used |
CE is the audiobox:CE prediction of frozen Humaneness Ears Medium. VoiceCLAP compares the generated audio to performance directions only: GENERAL plus round-bracket instructions, with the spoken SCRIPT text stripped from the reward text. With a reference, identity is an Orange WavLM-ID 250-dimensional cosine, not the separate 128-dimensional timbre metric. VoiceCLAP's reward is text-based even with a reference; the reference is not used as a performance target.
Each active reward is centered and standardized separately within its eight-candidate group, with standard-deviation floors and ±3 clipping. The weighted combination is centered and standardized again. The loss combines the custom VESPO sequence-importance weighting implementation with GDPO-style multi-component normalization, sampled KL coefficient 0.02 and SFT target anchor 0.02. This experimental combination is not a claim of superiority over ordinary GRPO.
Method references: VESPO: Variational Sequence-Level Soft Policy Optimization and GDPO: Group reward-Decoupled Normalization. Their published experiments concern language-model tasks; adapting them to this hierarchical speech architecture is our experiment, not a result established by those papers.
Candidates need finite waveform, genuine end-of-sequence and Parakeet word error rate ≤15%. A group requires at least four eligible candidates. If either member of a paired reference/no-reference batch fails, both groups are logged but excluded from optimizer updates. The run stops if paired rejection exceeds 20% after its minimum observation window; invalid audio is never given a positive reward merely for sounding pleasing.
Dataset and actual completed training
The input has 2,847 training cases: 1,362 synthetic voice-profile cases and 1,485 Emilia-provenance ladder cases; German 1,327, English 1,520. Each UID has a different-transcript 5.12–14.96-second speaker reference. Profile IDs and audited speaker-similarity checks govern synthetic matching. Explicitly robotic/silent/multi-speaker descriptions were excluded; provenance alone does not prove every Emilia clip is human-only. Small also filters short text and implausible word rates. The two models therefore do not have identical input pools.
Two G8 cohorts ran in parallel on 16 GH200 GPUs / four JUPITER nodes: one reference cohort and one no-reference cohort use the same UID. Groups are not individual samples: eight candidate recordings belong to each group, and two optimizer updates are made for each accepted paired batch. The completed run attempted 2,616 groups, trained on 2,434, rejected 182, and performed 2,434 optimizer updates. Training loop time was 3.83 hours, excluding allocation/startup tests. Final groups were budgeted from measured throughput, not silently expanded to the initially requested 10,000.
Peak LR 1e-5, 5% warmup then one cosine decay to 10% of peak; rank 128, alpha 128, residual scaling one. Adapters cover semantic q_proj/k_proj/v_proj/o_proj and local Talker attention c_attn/c_proj. Trainable adapter parameters: 38,666,240. Checkpoints were saved every 50 optimizer updates or 15 minutes with AdamW states and all rank RNG states. This public export contains adapter tensors, not the local optimizer/RNG checkpoint. A new run may initialize from this adapter, but exact resume requires the separately retained full training.pt and rank RNG records. Training and reproduction details.
Matched evaluation: improvements and trade-offs
There are 207 identical prompt/seed/reference cases per model: 121 reference, 51 name/alias, and 35 genuinely unconditioned/no-name cases. Comparator “matched base” uses the same FP32-rounded-base path and 0.7 audio temperature as the adapter. This is preferable to attributing differences from the usual BF16/temperature-1 Alpha column solely to RL.
| Condition | Model | Cases | CE ↑ | Genuineness ↑ | WER ↓ | Identity cosine ↑ | VoiceCLAP text cosine ↑ |
|---|---|---|---|---|---|---|---|
| reference | Default Alpha (BF16, temp 1) | 121 | 6.1785 | 4.3824 | 3.81% | 0.3877 | 0.0162 |
| reference | Matched base (FP32, temp .7) | 121 | 6.0328 | 4.3509 | 5.86% | 0.4110 | 0.0029 |
| reference | RL v1 (FP32, temp .7) | 121 | 6.4102 | 4.4657 | 2.13% | 0.4138 | 0.0088 |
| alias | Default Alpha (BF16, temp 1) | 51 | 6.1217 | 4.4394 | 4.53% | 0.2978 | -0.0690 |
| alias | Matched base (FP32, temp .7) | 51 | 6.0193 | 4.3684 | 5.07% | 0.3014 | -0.0608 |
| alias | RL v1 (FP32, temp .7) | 51 | 6.4109 | 4.5620 | 6.04% | 0.2989 | -0.1092 |
| none | Default Alpha (BF16, temp 1) | 35 | 6.1506 | 4.3277 | 1.64% | not applicable | 0.0753 |
| none | Matched base (FP32, temp .7) | 35 | 6.0692 | 4.2864 | 2.46% | not applicable | 0.0622 |
| none | RL v1 (FP32, temp .7) | 35 | 6.4345 | 4.4455 | 2.30% | not applicable | 0.0362 |
Against the matched FP32/.7 control, CE rises in all three conditions, reference-conditioned WER improves and speaker identity is approximately unchanged; this is not a clear voice-cloning breakthrough. VoiceCLAP instruction similarity falls in several conditions, despite being in the reward. Against the usual BF16/temperature-1 Alpha default, the WER comparison is not uniformly favorable—for example Large reference WER is worse after RL. Both baselines are therefore shown rather than treating a precision/sampling change as an adapter improvement. Genuineness is shown as an evaluation metric, not as a directly optimized v1 component. All judge outputs are model predictions, not listener MOS or identity certificates. Alias results are not representative of the RL training distribution. Some reference voices overlap training; this is a controlled listening comparison, not a speaker-disjoint benchmark. Full mean, standard deviation, counts and prompts are in stats/matched_evaluation.json and stats/evaluation_manifest.json.
License and responsible use
New adapter weights and authored documentation are released under CC BY 4.0 with attribution to LAION contributors. Upstream base/code/codec and judge dependencies keep their own licenses; no Orange judge weights are redistributed or relicensed. See NOTICE.md. Research and dataset augmentation are recommended uses. Obtain consent for voice cloning, disclose synthetic speech and do not impersonate people, mislead listeners or use this as an authentication mechanism. Production deployments need independent quality, bias, security and consent review. LAION does not endorse misuse.
Model tree for laion/Humaneness-Voice-Small-1.1-Alpha-RL-r128-v1
Base model
laion/Humaneness-Voice-Small