paulalesius/terence-mckenna-transcripts

Terence McKenna Transcripts

Catalog of machine-transcribed talks and interviews, mostly by Terence McKenna, built from 172 videos. Two tables:

• talks — one row per video; full original transcript (including [SPEAKER_XX] diarization tags)
• turns — one row per non-empty line; speaker tags stripped from text

Speaker labels are not consistent across files (no cross-video voice fingerprinting), so turns has no speaker column.

Load

from datasets import load_dataset

talks = load_dataset("paulalesius/terence-mckenna-transcripts", "talks")
turns = load_dataset("paulalesius/terence-mckenna-transcripts", "turns")

Configs

talks

column	type	description
id	string	Video title slug (filename without .transcript.txt)
title	string	Same as id, lightly cleaned
filename	string	Original source filename
transcript	string	Full file text, including speaker tags and line breaks
num_chars	int	len(transcript)
num_turns	int	Number of turns produced from this file

172 rows.

turns

column	type	description
id	string	Same id as the parent talk
title	string	Same title as the parent talk
turn_index	int	0-based line index within that talk
text	string	One transcript line, [SPEAKER_XX] prefix removed

135,672 rows.

A line like:

[SPEAKER_03] Why are we willing to go along with this shell game?

is stored in turns as text = "Why are we willing to go along with this shell game?" and remains tagged in talks.transcript.

Size (from the current build)

Talks	172
Turns	135,672
Characters in talks.transcript	14,068,290
Chars / talk	min 821, median ~60k, max ~476k
Turns / talk	min 10, median ~549, max 5,000

Long workshops dominate the upper end (e.g. Search For The Original Tree Of Knowledge, Esalen residencies). A handful of files are very short clips or fragments.

Source and processing

• Source: 172 video recordings; each transcript file was named after the video.
• ASR + local speaker diarization produced lines of the form [SPEAKER_NN] utterance.
• SPEAKER_NN is per file only. The same label in two files is not the same person.
• Empty lines and tag-only lines were dropped from turns.
• No cross-video deduplication was applied. Some titles look like the same talk uploaded twice (e.g. overlapping “Tryptamine Hallucinogens”, “Definitive UFO Tape”, “New & Old Maps of Hyperspace”, “Rap Dancing” recordings).
• The catalog includes some Dennis McKenna talks and multi-speaker trialogues (McKenna / Sheldrake / Abraham), not only Terence solo lectures.

ASR errors are present (misheard names, broken sentences, missing punctuation).

Intended use

Research, search, citation, language modeling, and retrieval over this lecture archive.

Not intended as a verified scholarly edition of McKenna’s words.

Limitations

• Automatic speech recognition, not human-corrected transcripts.
• No timestamps.
• No global speaker identities.
• Possible duplicate recordings.
• Mix of Terence McKenna, Dennis McKenna, interviewers, and panelists.
• Very short files may be trailers, readings, or incomplete jobs.

License and copyright

This repo packages unofficial machine transcripts I made from publicly circulated videos.

• I am not the copyright holder of the talks, interviews, or readings.
• I am not affiliated with the McKenna estate, publishers, or original venues.
• I do not grant a license to Terence McKenna’s (or Dennis McKenna’s, or any guest’s) words.
• What I did — diarized ASR, filenames aligned to the videos, and this table layout — you may reuse for research and model work.
• [SPEAKER_XX] labels are local to each file. They are not voice IDs and do not identify anyone across talks.

If a rights holder objects, the files should come down. If you use the corpus, you are responsible for how you use it. This is a research convenience, not an official text.

Citation

@misc{mckenna_transcripts,
  title  = {Terence McKenna Transcripts},
  author = {Paul Alesius},
  year   = {2026},
  url    = {https://huggingface.co/datasets/paulalesius/terence-mckenna-transcripts}
}