Release 02 · Research tools ·
Introducing Humaneness PerformanceNet: an open research library connecting authored intention with the careful description of what we see and hear.
“Keep the channel open. I’m still here.”
Why we built it
A scene has circumstances. A character has a history. An utterance has beats, pauses, physical effort, and a changing relationship to the listener. Describing a performance requires a vocabulary for more than an emotion label.
Humaneness PerformanceNet brings together script analysis, performance annotation, and complementary reference systems. It is a framework and research library for structuring that work, with browser-based reference materials and prompt-drafting tools.
It can support the design of synthetic performances, the organization of datasets, and task-specific supervision. It does not itself generate speech or provide a trained reward model.
An interactive scene study
Follow a fictional radio operator through five steps. This is an illustration of the workflow: the directions are authored examples, not measured model predictions.
01 · Character
Script analysis starts with the scene and the character’s makeup: physical, emotional, intellectual, and psychological context. These are authored premises, not a diagnosis of a real speaker.
Our fictional example: a tired radio operator keeps a rescue channel open.
02 · Intention
The script-analysis system uses seven Dramatic Obligations, interpreted through the character. Emotional Obligation is one part of that structure, connecting the scene’s demands to the performance.
For this example, the intended movement is from strain toward reassurance.
03 · Beats
Ordered Acting Beats carry words, actions, silence, and delivery. Each beat contains one nested Experience Chunk: the authored inner stream that gives the local action context.
The inner stream is a writing device. It is not evidence of a model’s or speaker’s private mental state.
04 · Vocabulary
Situation, emotion, voice, and visible or audible behavior each need their own vocabulary. The four references support structured prompts and annotations without collapsing everything into a single quality score.
A requested performance and an observer’s interpretation belong in separate fields.
05 · Evidence
The annotation system starts from recorded media and the modalities available. Time-linked observations stay distinct from qualified interpretations, observer impressions, and judgments about task fit.
Without an actual recording, observable fields remain unfilled. A prompt is not proof of what the output did.
Two connected workflows
Start from a book, screenplay, stage scene, radio script, or other source. Character Makeup and seven Dramatic Obligations provide context for ordered Acting Beats. One Experience Chunk inside each beat records its authored inner stream.
The result can inform directions for an audio or audiovisual generator: what happens, how delivery changes, and how each moment relates to the scene.
Start from the recorded performance. Describe the available modalities and time-linked events, then distinguish direct observations from qualified interpretations, impressions, and task-fit judgments.
A requested emotion is not automatically an observed one. Missing evidence stays missing, and an expression does not establish a person’s private psychological state.
The reference library
01
The conditions, tasks, and constraints around a performance. A reassuring line during a rescue has a different context from the same words in a rehearsal.
02 · 40 labels · 320 cells
40 emotion and state labels, organized into 320 intensity–expression cells. Authored feeling, requested expression, and observer interpretation remain distinct.
03 · 57 fields
A specialist vocal codebook with 57 fields. Its descriptors give speech a more precise vocabulary, with field-specific scales rather than one universal rating.
04 · 82 fields + 8 impression scales
82 descriptive fields: 79 ordinal fields and three categorical capture fields, plus eight separate observer-impression scales.
These counts describe the reference systems. They are not model capabilities, benchmark results, or a measure of dataset coverage. Each mark above stands for one field or cell.
Putting it to work
Use the library to encode scenes consistently, design varied synthetic performances, filter examples by explicit criteria, or group datasets by situation and behavior.
For training and evaluation, its structured descriptions can help define task-conditioned supervision. The judgment still depends on the task, the evidence, and how annotations are collected; there is no universal “good performance” score.
Next, we will share work on the speech side of this research direction: literal words, global scene intention, and local performance beats.
Open research library
The Hugging Face Space includes the integrated reference, Script Analysis System, Performance Annotation System, and supporting vocabularies. Consult the individual materials for their attribution and licensing terms.