Humaneness-Voice-Large-1.7

Humaneness Voice Large 1.7#

Expressive speech in English, German, French, Spanish and Dutch.

Write the words and direct the performance: emotion, conversational delivery, pauses and vocal reactions. Use a clean reference recording to guide the speaker identity, or generate a voice without a reference.

▶ Watch the video: Humaneness Voice Large 1.7

Optimized streaming and batch serving#

Large 1.7 now has a tested BF16 serving path with cached references, semantic CUDA graphs/Inductor compilation, optimized acoustic-token and codec decoding, and growing streaming chunks. The portable infer.py below remains a separate full-utterance reference path; serving flags do not automatically modify it.

These are matched single-stream results on one RTX 3090: CFG 2, temperature 0.8, the same structured prompt/reference and three warm repeats, with final alignment disabled. First PCM is a local HTTP measurement; RTF is TTS-plus-codec device time divided by generated duration. Browser playback, ASR, LLM and cold loading are separate. Large reaches 15.45 audio seconds/s in a short 16-candidate offline probe at CFG 2; the deployed batch cap remains 8.

Measured optimization recipes, Small/Large tables, hardware and reproducible code · Inference and batching notes · Switch models in Robby. Historical Small measurements on Jetson Thor are labelled separately in the report.

Listen and find your prompting recipe#

The manual covers 455 target/language/condition combinations. Its main audio players are visible directly; starting a clip pauses the others. Expand a condition for the remaining test takes, exact instructions and CFG comparison. Exploratory results are labelled, so an audible example is not automatically a validated recommendation.

Prompting in one minute#

Pass two separate inputs: text is exactly what should be spoken; prompt describes how to perform it. The standard format is:

GENERAL: A natural, close conversation. The speaker is trying to contain
strong anger rather than turning the whole line into a shout.
SCRIPT:
(intensely angry, held back) [3.2 seconds duration] You knew what this meant to me.
[0.4 seconds pause]
(quieter, the anger still present) [2.5 seconds duration] And you did it anyway.

The separate text is:

You knew what this meant to me. And you did it anyway.

• GENERAL: describes the overall voice, scene and audible delivery.
• SCRIPT: repeats the spoken words and places directions at the relevant beat.
• Round parentheses contain local delivery or non-speech reactions, such as (quietly angry) or (sharp gasp, 0.3 seconds). Keep these out of text.
• Square brackets can request duration and pauses. Some searched recipes instead place duration inside round parentheses; copy the recorded recipe's syntax when reproducing its results.
• Keep performance instructions in English where practical; dialogue can be in any of the five supported languages.

Requested timing is learned conditioning, not a guaranteed editing boundary. For masking, hesitation, vocal bursts and complete examples, read the prompting guide. To choose wording and CFG for a particular emotion or style, use the interactive manual.

Quick start#

This repository contains complete merged 1.7 weights. Ordinary inference does not require attaching an adapter.

Install a CUDA-enabled PyTorch build for your server, then the dependencies:

hf download laion/Humaneness-Voice-Large-1.7 --local-dir humaneness-large17
cd humaneness-large17
pip install -r requirements.txt
python infer.py --language en --seconds 9.2 --seed 1701 \
  --text 'I thought you had forgotten. You remembered every little detail. I am so glad you came back.' \
  --prompt-file examples/prompt-en.txt --output example.wav

The custom implementation requires trust_remote_code=True; review the code before loading. Root weights are stored in FP32. The supplied inference loader uses BF16 weights and compute. The separate MOSS Audio Tokenizer v2 retains its upstream terms.

Add a reference voice#

Add --reference-wav clean-speaker.wav to the command above. Use clean speech from one speaker, with different words from the requested output. Supply at least 5.04 seconds; the helper crops the waveform to 14.96 seconds before encoding. It does not use the historical 2.96-second cap. Reference audio guides identity but does not guarantee exact voice preservation.

For repeated requests, keep the model and codec resident:

from pathlib import Path
import infer

model, processor, api = infer.load()
api.load_codec(processor, "OpenMOSS-Team/MOSS-Audio-Tokenizer-v2")
reference = infer.reference_codes(processor, Path("clean-speaker.wav"))

codes = infer.generate_codes(
    model, processor, api,
    text="I am so glad you came back.",
    prompt="GENERAL: Warm, natural relief.\nSCRIPT:\n"
           "(quiet relief) [3.0 seconds duration] I am so glad you came back.",
    seconds=3.0, language="en", seed=1701, reference=reference,
)
waveform, sample_rate = api.decode(processor, codes)

Adjust classifier-free guidance#

For a guided take, try:

python infer.py --language en --seconds 9.2 --seed 1701 \
  --text 'I thought you had forgotten. You remembered every little detail. I am so glad you came back.' \
  --prompt-file examples/prompt-en.txt \
  --cfg 2.5 --neutral-prompt-file examples/neutral-en.txt \
  --output example-cfg.wav

The neutral branch must keep the same literal words, durations and reference; remove only the emotional/style treatment. CFG 1 means guidance off. CFG 2.5 is the acting showcase setting, not an optimum for every condition. The manual reports condition-specific CFG searches. Those searches used the merged 1.7 model without reference audio or extra adapters; preserve that setup when reproducing them.

Use the manual as a person or an agent#

Choose a language and an emotion, talking style or vocal-burst condition in the interactive manual. Listen to the selected recipe and paired baseline, then inspect the exact conditional and neutral prompts. All six held-out takes remain available, including failures; recipe selection does not cherry-pick a test recording.

For programmatic use, start with manual/AGENT_GUIDE.md and manual/recipes.jsonl. Match the target, language and, for vocal bursts, surrounding-emotion condition. Check usable and the evidence status before treating a recipe as a recommendation. The linked evidence preserves prompts, sampling settings, raw scores and audio.

The earlier 1.5/1.6 manuals document different weights. Their specialist adapters and preferred CFG values are historical evidence; use the 1.7 manual above for this model's measured recipes.

What is included#

Do not apply the included adapter to the merged root model again. For the original unmerged arithmetic, python infer.py --unmerged ... loads parent_model/ plus the adapter exactly once. This native control is useful because merging changes BF16 rounding and can change sampled speech.

Training and model lineage#

1.7 combines the Acting High-8654 supervised checkpoint with its selected rank-32 acting adapter at update 44,250, scale 1.0. This is a supervised branch from Large 1.3 Alpha. It does not contain the 1.4/1.5 RL adapter stack and is not Large 1.6 with another adapter attached.

The 1.3 ancestor first received a full supervised acting continuation. The High-8654 checkpoint then received a rank-32 adapter trained on a per-language union of top 1% emotion/style takes and top 2% genuineness-plus-blend takes. The adapter schedule ran for three epochs; 1.7 selects update 44,250 rather than the final 58,944. Training covers EN, DE, FR, ES and NL. See TRAINING.md for source mixture weights and the full recipe.

To fine-tune with compatible data, start with training/train_reference.py.

Architecture#

The hierarchical MOSS Local Transformer TTS model has 4,129,990,144 unique base parameters: a Qwen3-family semantic transformer and a local Talker. Each 80-ms frame produces 12 indices, one per codec codebook, each with 1,024 possible values: 12.5 frames/s × 12 = 150 codec indices/s. The separate codec returns 48-kHz mono audio. The model does not generate all 32 codec books.

Checkpoint selection results#

Seven second-half adapter checkpoints were compared on the same 108 English/ German tasks: 40 emotions and 14 edge cases, one seed, no reference, CFG 1. Update 44,250 had the highest frozen composite among the seven complete matrices:

The target diagnostic is not an accuracy percentage. WER is the arithmetic mean of per-clip rates, not a corpus-token-weighted rate. Scores are model predictions, not human listening ratings. The selected state does not maximize every metric. This was development selection among seven checkpoints, not an independent held-out confirmation. See stats/selection.json for the selection record and paired uncertainty.

Those selection recordings use the original unmerged BF16-parent/FP32-LoRA path. FP32 merging was checked algebraically. The merged BF16 deployment was then tested against that native control on one fixed script in each language and one English CFG 2.5 pair. Every take is retained. The merge audit does not claim bit-identical sampled tokens or establish quality gains from five scripts. You can also hear the canary audio or download its listening HTML.

The separate 1.7 evolutionary manual tests recipes on the merged model; its held-out recipe evidence should not be confused with the checkpoint selection study above.

Throughput and streaming#

Reuse a resident model and codec, cache reference codes and batch compatible candidate generations. Best-of-N selection must finish its candidates before choosing a winner. Streaming latency and offline throughput are separate goals.

The Humaneness-Voice server contains batching, streaming and optional graph/kernel optimizations. Its published 0.16-second onset is a Small-model Jetson Thor result, not a 1.7 measurement. This release provides a full-utterance reference path and batched same-condition CFG helper; it does not claim that every historical server or CUDA graph is validated for 1.7. Read the inference and optimization guide before reporting speed measurements.

License and use#

New model contributions and release materials are CC BY 4.0. Keep NOTICE.md and upstream notices. External codecs, software and training datasets retain their own terms. CC BY does not grant rights over another person's identity. Use consented references, identify synthetic material appropriately and test the model before deployment.