LAION’s Humaneness Voice Converter takes an existing spoken recording and attempts to reproduce its content in another consenting speaker’s voice, while retaining much of the original delivery. The release includes a complete inference checkpoint and a local command-line workflow. Users do not have to merge an adapter into another model.
LAION’s evaluation shows a trade-off. Compared with untouched Chatterbox, the converter scores better on predicted content enjoyment, predicted production quality, and source-emotion matching. It scores worse on target-speaker similarity. A rendition that a scoring model prefers may not sound most like the intended speaker, and neither result establishes that the words survived unchanged.
The Startrise release tracker lists the release on October 11, 2026. LAION’s model card supplies the substantive evidence: the runnable package, its evaluation conditions, and limitations that should accompany any local experiment.
A Complete Speech-to-Speech Checkpoint, Not a TTS Model
Humaneness Voice Converter is based on Resemble AI’s Chatterbox, with the voice-conversion architecture and training developed by LAION. It is an audio-to-audio system, not an LLM or a text-to-speech model, and is distinct from Humaneness Voice Small semantic TTS.
The source recording supplies speech content, timing, and performance. The target recording supplies the voice reference, documented as a clean, single-speaker clip lasting five to ten seconds.
The package includes the Chatterbox S3 speech-token extractor, CAMPPlus speaker encoder, score-conditioned flow decoder, HiFT vocoder, and frozen Humaneness Ears Medium score predictor. Users do not need to locate and combine several LoRA adapters before inference.
Chatterbox extracts roughly 25 semantic speech tokens per second from the source audio. The target supplies both a speaker embedding and a short acoustic prefix. Humaneness Ears Medium adds 99 conditioning scores, with time-varying emotion and style information drawn from the source and selected relatively stable voice attributes drawn from the target.
The decoder generates mel features, which represent the audio’s spectral content, and HiFT renders a 24 kHz waveform. Both target-conditioning inputs matter; this is not voice cloning from a text description alone.
How to Run Humaneness Voice Converter Locally
LAION’s local setup instructions assume a CUDA-enabled PyTorch installation. The reported test environment used Python 3.13.5, PyTorch 2.9.1+cu128, Chatterbox 0.1.7, and Transformers 5.14.1.
The commands were validated inside an existing Jupiter environment that overrides several upstream Chatterbox dependency pins. LAION does not present that arrangement as a fresh, cross-platform dependency-resolution test. The version list documents a working environment; it does not guarantee that installation will succeed unchanged on every machine.
With platform-compatible CUDA PyTorch installed and the hf command available, download the repository first so its requirements file is accessible:
hf download laion/Humaneness-Voice-Converter \
--local-dir humaneness-vc
python -m pip install -r humaneness-vc/requirements.txt
python -m pip install --no-deps chatterbox-tts==0.1.7
The --no-deps installation follows LAION’s documented arrangement. It does not resolve dependency conflicts automatically, so checking the resulting package versions is part of setup.
For one conversion, provide the source speech and an authorized target recording:
python humaneness-vc/infer.py \
--model-dir humaneness-vc \
--source source_speech.wav \
--target consenting_target_voice.wav \
--output converted.wav \
--seed 20261011 \
--flow-steps 10
The validated input path accepts finite audio samples lasting one to 30 seconds, either mono or mixed down to mono. WAV works; other formats depend on the installed SoundFile build. For results comparable to the documented setup, use a clean five-to-ten-second target with one speaker.
LAION reports running the command on a GPU with a 6.528-second source and a 5.448-second target, producing a 6.52-second, 24 kHz WAV. This establishes a concrete tested path, without demonstrating performance across all recording conditions.
A fixed seed makes sampling repeatable on a fixed software and hardware stack. Different seeds produce different takes. Optional candidate selection can generate and rank several:
python humaneness-vc/infer.py \
--model-dir humaneness-vc \
--source source_speech.wav \
--target consenting_target_voice.wav \
--output selected.wav \
--candidates 8 \
--microbatch 8 \
--precision bf16 \
--metrics selected.metrics.json
Candidate selection is a more expensive inference strategy and is not the default. With one candidate, the script does not run candidate ranking. The ranker combines standardized predicted enjoyment, target-speaker cosine similarity, and negative emotion error with equal weights. Its judgment remains an automated proxy, not a listener’s assessment.
Predicted Quality Improves While Speaker Similarity Falls
LAION’s reported comparison uses 40 curated source recordings and two fixed target voices. The default results are single-sample outputs; best-of-16 results add candidate selection.
| Decoder | Predicted Content Enjoyment ↑ | Predicted Production Quality ↑ | Emotion MSE ↓ | Target-Speaker Cosine ↑ |
|---|---|---|---|---|
| Untouched Chatterbox S3Gen | 6.169 | 7.766 | 0.113 | 0.821 |
| Humaneness Voice Converter | 6.453 | 8.009 | 0.092 | 0.773 |
| Humaneness Voice Converter, best-of-16 | 6.460 | 8.009 | 0.072 | 0.795 |
The default converter gains 0.284 points in predicted content enjoyment and 0.243 in predicted production quality. Emotion mean squared error falls from 0.113 to 0.092, indicating closer matching on the source’s two strongest emotion axes under this evaluator.
Target-speaker cosine similarity falls from 0.821 to 0.773. Best-of-16 selection recovers some of that loss, reaching 0.795, but remains below untouched Chatterbox’s 0.821. Predicted enjoyment barely changes with selection, and the reported production-quality score stays the same. Candidate selection therefore looks more plausible as a tool for improving emotion matching and speaker consistency than for substantially raising these quality scores.
Cosine similarity measures similarity between speaker embeddings; it is not a percentage of voice-cloning accuracy. A lower score is relevant evidence about target matching, but it cannot identify every audible identity difference or verify who is speaking.
The Evaluation Is Not an Independent Listening Study
Content enjoyment, production quality, and emotion scores come from Humaneness Ears Medium, the same judge family used during training. These quality figures need careful reading.
LAION’s later full-flow training weighted predicted content enjoyment at 40%, source-emotion matching at 30%, and target-speaker similarity at 30%. Those weights help explain the model’s objectives, but they do not establish that listeners prefer the resulting compromise. Evaluating with a related training judge is not automatically invalid; improvements should be read as progress against that scoring system until independent evidence establishes broader benefits.
Sources
- Humaneness Voice Converterhuggingface.co
- Startrise release trackerstartrise.io
