LAION opens Humaneness Voice 1.1 Alpha checkpoints while training continues
Large and small speech models, interim weights and a public listening comparison are available now. LAION says training and quality evaluation are still in progress.
Listen before the training run ends
LAION has published Humaneness Voice 1.1 Alpha as large and small experimental text-to-speech checkpoints. The October 13 Hugging Face repositories already contain complete weight files for interim training steps, along with inference code and configuration. An English listening page compares checkpoints across several delivery styles and reference-voice cases, with prompts and audio also available in a public dataset. The models synthesize literal text with optional directions for emotion and pacing, and can use a short reference recording or a trained alias for voice identity. These are continuations of earlier models, not training runs started from scratch.
What is still unproven
Both model cards state that training is still running. LAION explicitly does not claim an improvement in speech quality before evaluation completes, and the listening page labels pending comparisons. The large card describes a tested GH200, PyTorch and Transformers setup, while warning that other stacks need portability testing. Its published research guidance asks users to obtain speaker consent, disclose synthetic audio and test identity and language behavior before deployment. AITrending has verified that the repositories expose interim weights and a public demo, but has not assessed the generated audio or benchmarked either model.