A public buyer pays hundreds of times more per hour to have specialised speech made than the same data later lists for in a research catalogue
NIST paid the University of Pennsylvania $645,812.76 for the SAFE-T public-safety speech corpus, specified at a minimum of 291 hours (over 300 delivered, 122 or more transcribed): about $2,150 to $2,220 an hour, all-in. LDC now lists the 141-hour OpenSAT 2019 portion at $1,000 for a non-member research licence, about $7 an hour, roughly 300 times less. The same gap shows in Spain: the Barcelona Supercomputing Center paid about EUR 1.0 to 1.6 per Catalan example written and annotated to order (NLUCat, InToxiCat, CoQCat), but about EUR 0.012 per sentence pair for 15 million Catalan-English pairs a vendor already held.
Dated 15 Aug 2023Source checked 27 Oct 2026Quote found on its page again 28 Oct 2026Basis: indicative
Selling a company's data? The guide beside the button shows what this record means for a sale.
- No data upload
- About two minutes
- No email needed
“"description":"PUBLIC SAFETY SPEECH CORPUS FOR SPEECH ANALYTIC TECHNOLOGY (OPENSAT) EVALUATIONS","total_obligation":645812.76”
USAspending.gov, NIST contract DOCSB134118CN0003 (public safety speech corpus), checked 28 Dec 2017
Delgado et al., The SAFE-T Corpus (LREC 2020), checked May 2020
LDC catalog, 2019 OpenSAT Public Safety Communications Simulation (LDC2023S06), checked 15 Aug 2023
TED contract award notice 2022/S 250-731229, Barcelona Supercomputing Center, checked 28 Dec 2022
- SAFE-T public-safety speech corpus made to order: at least 291 hours of game-elicited English speech under loud background noise from 100 or more speakers, at least 122 hours manually transcribed, to train and test speech recognition (NIST OpenSAT, built by the Linguistic Data Consortium)
- 2019 OpenSAT Public Safety Communications Simulation (LDC2023S06): about 141 hours of noisy, game-elicited US English speech with transcripts, a portion of the NIST-funded SAFE-T corpus
- Segmentation, alignment and translation of a speech corpus, to order, for language-model training (Barcelona Supercomputing Center, Aina project, lot 9)
- Translation of multilingual reference corpora (benchmarks) into Catalan for evaluating language models (Barcelona Supercomputing Center, Aina project, lot 1)
- CA-EN Parallel Corpus: about 15.0 million Catalan-English sentence pairs, human or human-proofread translation, tagged by domain and text type, bought with translation from a language vendor's own database under CC-BY (Barcelona Supercomputing Center, Aina project)
- caBreu: 3,000 Catalan news articles (about 700 words each) with extreme, abstractive and extractive summaries written by three annotators (Barcelona Supercomputing Center, Aina project)
It is shown on its own, not counted into a figure.
Cited in this guide