~/projects/byu-speeches/methods
Methods and data
Where every speech comes from, what is counted and how, what is missing, what this site publishes and what it keeps offline, and a card for each study as it ships.
Sources
| Years | Source | Count | Notes |
|---|---|---|---|
| 1891– | speeches.byu.edu, the public WordPress API | 2,546 speeches | Transcript, title and BYU's event and topic tags. |
| 1891– | Each speech's page | 2,255 with audio | Delivery date, the speaker's position line, the audio file and the video. |
Speeches by decade
- 1890s2
- 1930s2
- 1940s6
- 1950s181
- 1960s154
- 1970s290
- 1980s305
- 1990s337
- 2000s492
- 2010s471
- 2020s306
By calling when speaking
- Quorum of the Twelve 548
- Seventy 459
- BYU leadership 370
- BYU faculty 262
- Other 125
- First Presidency 107
- Presiding Bishopric 42
- Relief Society 30
- Young Women 18
- Sunday School 12
- Primary 12
- Young Men 8
From the role line under the title where the source prints one. Where the source gives only a title such as Elder or President, the calling is the one the leadership database records for that speaker in that year.
What is measured
- Words
- The prose paragraphs of the transcript. Headings, the notes after a transcript and the byline are excluded. A speech whose page carries under a hundred words of text is treated as having no transcript (557 speeches, mostly recordings from the 1950s and 1960s).
- Reading grade
- Flesch–Kincaid grade level on the prose, with syllables counted as vowel groups. Nothing under fifty words is graded.
- Scriptures cited
- References written out in the text and in the notes that follow it. Only references with a chapter and verse count, so a bare book name is not a citation.
- Speakers
- Matched across sources by full name with suffixes kept, so Brigham Young and Brigham Young Jr. are different people, and split by year where two leaders share a name. A short alias table resolves byline variants. Speakers are linked to the leadership database by the same key.
- Signature phrases
- Words and two-word phrases scored by weighted log-odds with an informative Dirichlet prior against the rest of the corpus, after removing stop words, names of deity and the ordinary vocabulary of a sermon.
- Hours of audio
- Estimated from each recording's file size and bitrate without downloading it, exact for constant-bitrate files.
What is published and what stays offline
The speeches are copyrighted by Brigham Young University. The full texts are stored offline for analysis and never enter the site. The site publishes metadata, aggregate measures and a single opening sentence per speech, with a link to the official page for everything else. Models built on the text ship as their outputs or as small browser-side models, never as the text itself.
Model cards
One card per study, added as each study ships: data, method, validation and known failure modes.
- Guess the year supervisedModel card
- How phrases spread unsupervisedModel card
- Who quotes whom unsupervisedModel card
- Who gets quoted supervisedModel card
- What a paragraph does supervisedModel card
- Who the talk is for supervisedModel card
- What rose and what faded unsupervisedModel card
- How words changed meaning unsupervisedModel card
- Style fingerprints unsupervisedModel card
- The stories they tell unsupervisedplanned
- The shape of a talk unsupervisedModel card
- Whose talk is this supervisedModel card
- What outperforms the speaker supervisedModel card
- What makes a hall laugh supervisedplanned
- How they sound unsupervisedplanned
- What only Latter-day Saints say unsupervisedModel card