Home About Publications Dataset Platform

Project Themes:

The Jerk-O-Meter

MoodPhones / VibePhones

Elevator Rater

Human Interest-Meter

Speed Dating v2

Negotiations

Movie Audience Reactions
GroupMedia

How Convincing Am I?

Project Lead: Will Stoltzman
Project Supervisor: Prof. Alex "Sandy" Pentland

The ElevatorRater is a speech analysis program which assesses charisma based on a speaker's delivery. The program looks only at non-linguistic speech features like pitch, speaking rate, pause durations, etc.

The ElevatorRater is so-named because one application is to evaluate the short, persuasive talks known as "elevator pitches". Everyone has an elevator pitch of their own (think about how you might introduce yourself and your interests to a new acquaintance), but these mini-speeches are particularly prevalent in the business and entrepreneurial worlds.


Sample output from the ElevatorRater in December 2005. The algorithm predicted a 5 on a scale of 1-10 for a particular speaker. A histogram of past speakers in that session was provided for comparison.

How the ElevatorRater works

We built our model of charisma and persuasion by running several identical experiments where participants were asked to give a short, prepared pitch on a topic of their choice, and then were immediately rated by their peers according to the following three questions:

  1. How persuasive is the speaker, apart from the details of the pitch?
  2. How convincing is the content of the pitch, apart from the way in which the speaker delivered it (e.g. if you had read it)?
  3. How effective is the presentation style, apart from the details of the pitch and the way in which the speaker delivered it? (where "presentation style" addresses number of "ums," senetence structure, pacing, flow of information, etc.)
The three questions (Q1-Q3) were designed to isolate different aspects of a pitch, but it turns out that these aspects are very closely related:
corr(Q1, Q2) = .82
corr(Q1, Q3) = .94
corr(Q2, Q3) = .77
[N = 42, all p-values < 10^-8]
This implies one of two things: style and content are intrinsically intertwined, or listeners are not good at segmenting these attributes. As such, the ElevatorRater focuses on predicting Q1, taking it as a strong indicator of overall reception of a speaker.

In its current incarnation, the ElevatorRater takes three features of a pitch into account:

  • Standard deviation of the spectral entropy.
  • Mean-scaled standard deviation of the energy.
  • Voicing rate. This is number of sonorant sounds (vowels) per unit time uttered by a speaker while he is speaking (i.e. pauses between phrases won't affect the voicing rate).