~/projects/byu-speeches/voice

How they sound

A speech is heard before it is read. This study measures the voice in a stratified sample of recordings, pace in words per minute of speaking time, pauses per minute and how long they run, the share of the recording that is speech, median pitch and pitch range, clusters the speeches by those six numbers, follows them across decades and between men and women, and asks whether the way a speaker sounds matches the way the same speaker writes.

Unsupervised58 speeches sampledAudioProsody ~/byu-speeches

What stands out

The sample sorts into low share speaking, low words per minute, high share speaking (13, 30, 15 speeches), named by the trait that sets each apart.
Median pace went from 176.2 words per minute in the 1970s to 188.75 in the 2020s, pauses from 15.75 to 15.89 per minute, and pitch range from 15.46 to 21.56 semitones.
The BYU records do not carry gender, so the comparison by gender is on the conference hub only; by ear the sample has both.
Voice clusters barely follow written-style clusters (adjusted Rand index -0.015); the strongest single link is pitch range (semitones) with sentence length at r = -0.32.
Bruce R. McConkie (1978) is the fastest in the sample at 261 words per minute; Jacob de Jager (1982) the slowest at 142.1.

Loading the study…

Model card

Data: 58 speeches with both audio and a full transcript, drawn with a fixed seed per decade (sample_audio.py); the official recordings are read once and only summary numbers and two low-resolution tracks are kept.
Measures: Audio resampled to 16 kHz. Energy in 10 ms frames; a frame is speech when it is above a threshold set between the recording's noise floor and its speech level; a pause is 0.35 s or more below it. Pitch by the YIN estimator on speech frames between 65 and 400 Hz; range is the 5th to 95th percentile in semitones. Pace is the transcript's word count over speaking time, so it is an articulation rate, not a rate over the whole recording.
Clustering and comparison: Six features standardised, k-means with three clusters and twenty restarts. The same speeches are clustered on the style study's stylometric vector, and the two labelings are compared with the adjusted Rand index; Pearson correlations between each voice measure and six style measures are listed.
Known failure modes: Old recordings have a high noise floor and sometimes music or an introduction, which the energy threshold handles imperfectly. Transcripts can differ from what was said, which moves the pace. The pitch estimator is confused by reverberation and by two voices. A sample of this size describes the sample; decade medians rest on about ten speeches each.

Methods and data for the corpus as a whole.