Human Vocality Primitives

Breath & Noise Airstreams

Controlled airflow textures including exhalations, fricative noise, and shaped airstream bursts.

FULL DATASET

Human Vocality Primitives

This dataset is currently in production. Preview audio and full specifications will be added as they become available.

This is a preview release provided via Hugging Face for evaluation purposes. Initial recording passes are available to assess capture quality, labeling structure, and dataset relevance. The full dataset will be delivered privately with complete Proteus provenance, integrity, and documentation upon licensing.

This dataset is complete and available for licensing. A public preview subset will be released via Hugging Face on a rolling basis.

Breath and noise airstreams are airflow-driven respiratory behaviors produced without vocal fold engagement, relying instead on controlled inhalation and exhalation, pressure modulation, and open vocal tract configuration to generate sound. In these techniques, acoustic energy is created exclusively through turbulent airflow rather than phonation or articulation, resulting in pure breath-noise output with no pitch, voicing, or linguistic shaping. Expressive control emerges through modulation of airflow intensity, effort, continuity, interruption, and transitions into and out of silence rather than changes in loudness, pitch, or articulatory gesture.

Acoustically, breath and noise airstreams occupy a broadband, noise-dominant timbral space characterized by diffuse spectral energy, intensity-dependent texture, and fine-grained temporal variation. Expressivity is conveyed through airflow pressure, dynamic contouring, interruption patterns, and the contrast between inhalation and exhalation rather than through vocal tract resonance shaping. These characteristics make breath and noise airstreams particularly well suited for modeling non-phonated respiratory sound, effort-aware systems, and airflow-centric acoustic analysis, where precise control of intensity dynamics, gating behavior, and repeatable breath gestures is essential.

Audio Demonstrations

A three-part listening benchmark: a mixed musical demo built from this dataset, the raw source clip, and an AI model’s attempt to reproduce the same prompt.

PRODUCED REFERENCE

A musical demonstration created by replacing a state-of-the-art AI-generated lead instrument with original source recordings from this dataset, then arranged and mastered to preserve musical context. This approach allows direct comparison between current-generation model output and real, rights-cleared acoustic source material.

RAW DATASET CLIP

Directly from the dataset: an isolated, unprocessed example of the source recording.

AI MODEL BENCHMARK

An unmodified output from a current-gen AI model given the same musical prompt. Included to illustrate where today’s systems still differ from real, recorded sources.

AI model approximations generated using publicly available state-of-the-art music generation systems.