0

tl;dr; I'm looking for a corpus (or multiple of them) of un-scripted English speech, transcribed to text and tagged by who said what. Ideally, I'd like hundreds or even thousands of distinct speakers.

I've found a few possibilities, but they have issues. Also, I don't have much of a data science background (I'm a dev tools programmer by trade) so I frankly have no idea where to even start looking beyond the blindingly obvious search terms, and those haven't returned anything very inviting.


I'm working on a literary project where I want to create dialog with recognizably distinct "voices". I don't want everyone's speech to sound like they are quoting me, but rather seem like actual different people. I've a few different ideas of how to approach this, but all of them require training data with rather specific properties.

  1. I need text, not audio.
  2. I need data source from speech, not written materiel and I'd strongly prefer non-scripted speech. (I know that for some people, myself included, the way they speak and write can be very different from each other).
  3. I need data tagged by who came up with the words. (The same person giving speeches from different speech writes, or a screenplay where everyone's dialog is written by one person would be undesirable.)
  4. I want a diverse set of speakers. (A data set where everyone in it works in the same industry or is talking about the same subject could be problematic.)

On the flip side, my "openness" constraints are rather lax; I'm not going to pay for data is about the only hard constraint. I'd like to be able to redistribute the source data, but as long as I can make the cooked data available to a private group I'm fine with most restrictions.


Most of the natural langue data sets I've found fail hard at the second point: most text data set seem to be from written work like questions on this site or reviews/comments on other web pages. Also the third point is problematic, data sets with speaker tagging don't necessarily tag it "correctly" for my needs.

Off hand the only source I've come up with that would already exist would be court transcripts. However it looks like they are often far from free and/or open, thought some some are free to download, but maybe not in the easiest format. Also, I suspect that such transcripts will have significant over representations of legal professionals and legal topics, likely more so at the higher court level which are the ones likely to be easiest to obtain.

1 Answer 1

0

I don't have a link to a specific dataset, but maybe some podcast transcripts would give you what you're looking for? I would also search for "dialogue dataset", here is e.g. one option: https://paperswithcode.com/dataset/dailydialog

2
  • Interesting data set. However I'm not seeing any talk of how it was generated. For example it could all be from a single person role-playing both sides of each conversation. Also the CC NC-SA licence could be a problem: at what point is my work no longer "building" on that data? Publishing a "is this that speaker model" would be fine, but I need to keep the copyright for the ultimate end product.
    – BCS
    Jan 23 at 21:51
  • Podcasts would be a possibility, but I'd need a large archive (so getting them peace-meal would be problematic) and I'd need it accurately speaker tagged (which I suspect isn't at the top of many people's priority list).
    – BCS
    Jan 23 at 21:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.