Speaking, Fast and Slow Competition - Development Data

Speaker, Fast and Slow Competition Development Data

This data contains read-speech recordings at both natural and deliberately-slowed speech from a speaker of American English and a speaker of British English. There is a total of 22 minutes of natural speech and 38 minutes of corresponding slowed speech.

The "Speaking, Fast and Slow" competition asks participants to produce slowed versions of natural speech in a way that sounds natural (i.e. how humans actually sound when speaking slowly). The main application of this technology is for language learning: a successful system would enable a language student to take any natural recorded speech and create a slow, but still natural-sounding version (the equivalent of having a language teacher who is able to speak more slowly/clearly to aid in comprehension), without some of the strange acoustic artifacts of typical straightforward methods of slowing speech.

Data Collection

The text corpus is a pool of CEFR-graded sentences, initially generated by an LLM and then verified and edited by human experts, These were then read aloud and recorded at a natural speed by the participants, followed by a reading of the same sentence with a longer target time (guided by the recording application). For the slow recordings, participants were instructed to "imagine you are speaking for an English language learner trying to help them better-understand the sentence, i.e. try not to speak louder necessarily, but enunciate clearly and more slowly."

Speakers

In total the development data has 2 speakers, both male, one of American English and one of British English.