The emotion model
Valence and tension — a reading of the room, not a guess
Prosody scores each line of a transcript with a RoBERTa model fine-tuned on EmoBank, a research corpus of text annotated for emotional dimensions. We only show what held up when we checked the model against human ratings of real meeting transcripts — not everything the model is technically capable of scoring.
Valence
positive ↔ negative
Whether the language is warm and constructive or frustrated and negative. A sustained dip is usually where a decision started to slip. This is the one score we show as a number — it held up when we checked it against human ratings of real meeting transcripts.
Tension
calm ↔ escalating
A ranking, not a raw score: the moments where negative valence lines up with a spike in the energy behind the words. That combination is what reliably marks a tense exchange, not just an animated one.
Mood timeline
The per-line scores roll up into a timeline for the meeting. You see the shape of the discussion at a glance and jump straight to any point in the transcript.
Tense-moment detection
Where negative valence lines up with a spike in the energy behind the words, Prosody marks a tense moment. Those stretches are where decisions tend to stall, and they feed the recurring-tension view in analytics.
A fine-tuned RoBERTa model
The scorer is a RoBERTa transformer fine-tuned on EmoBank — a corpus of sentences annotated by researchers on the classic Valence-Arousal-Dominance emotion model. Prosody runs it on its own infrastructure, so the emotional-tone scoring happens entirely inside Prosody.
Each line comes back with three continuous scores. Valence and the tension ranking built from it are what reach you — the drafting step reads them alongside the words so a blocker and an aside are not treated the same way.
Only what holds up
We didn't stop at the research benchmark. We checked the model against human ratings of real meeting transcripts, not just EmoBank's news-and-blog-post text it trained on. Valence held up. The tension ranking held up. The model's third dimension — dominance — didn't: on our own test set, two independent human raters actually disagreed with each other about which direction it even pointed. A score nobody agrees on the meaning of isn't a score, so we don't ship it, and we don't show arousal on its own as a number either — only in the tension ranking, where it earns its keep. We'd rather show you less and have it be right than hand you a dashboard full of numbers we can't stand behind.
A signal, not a verdict
Emotion scores are an input to a human decision, never a replacement for one. They point a reviewer at the moments worth a second look; they do not decide priority, they do not assign work, and they never create a ticket. A person still reads every draft and confirms it against a real account before anything is created.
Try it
A rough, in-browser feel for the two axes that matter most. Drag either slider and watch the trace reshape.
Drag to feel how valence and arousal reshape a moment.
positive ↔ negative — how warm or hostile the language is
calm ↔ activated — how much energy is behind the words
A toy for intuition only. In the product only valence is shown as a number; this second slider stands in for the energy behind the words, which feeds the tension ranking rather than a score of its own. A reviewer still confirms every draft.