The IFA corpus: a phonemically segmented dutch "open source" speech database

Abstract

An open source database of hand-segmented Dutch speech was constructed with off-the-shelf software using speech from 8 speakers in a variety of speaking styles. For a total of 50,000 words, speech acquisition and preparation took around 3 person-weeks per speaker. Hand segmentation took 1,000 hours of labeling altogether. The asymptotic segmentation speed was about one word, or four boundaries, per minute. An evaluation showed that the Median Absolute Difference of the segment boundaries was 6 ms between labelers, and 4 ms within labelers. Label differences (substitutions, insertions, and deletions) were found in 8% of the segments between labelers and 5% within labelers. Compiled data are available in relational database format for querying with SQL.

Real-time synthesis of imagined speech processes from minimally invasive recordings of neural activity

“…Each run consisted of 100 Dutch words, which were displayed for two seconds followed by a fixation cross for one second. Words were drawn randomly from a phonetically balanced list of 250 words 19 .…”

Adjustable deterministic pseudonymization of speech

“…Stimuli were Pseudosentences from the IFA corpus read by 10 Dutch speakers (5F) (Van Son et al, 2001). In Experiment 1, the parameters of the male-like Long VTL target are 𝜙 = 510 Hz, F 0 = 120 Hz, rate = 3.8 syll/s.…”

“…To pseudonymize the formants, the target frequencies, represented in terms of VTLs 𝜙 𝑖 = 𝐹 𝑖 ∕(2𝑖 − 1), can be randomly chosen in the range 𝜙 𝑖 ±40 and 𝜙 𝑖 ±75 Hz for F 0−1 and F 2−5 , respectively, and the intensities can be randomly chosen in the range 64 ± 4.5, 67 ± 2.5, 58 ± 4.5, 50±8, 47 ± 10, 45±9 dB (I 0−5 ±2SD), where SD denotes standard deviation. These values were chosen based on ranges found in the speakers in the IFA corpus (Van Son et al, 2001) (5M/5F, see Experiment 1, Section 3.1.1). In an alternative setting where a given speaker is to be pseudonymized to a specific target speaker, the parameters 𝜙, 𝜙 𝑖 (𝑖 = 1...5), 𝐼 𝑖 (𝑖 = 0...5) and speaking rate can be pre-computed across several of the target speaker's utterances (preferably over 300 s spoken in a comparable style) and used.…”

Acquiring and implementing phonetic knowledge

“…As two representative examples let me shortly describe the 10-millions words Spoken Dutch Corpus (CGN) [17] and the much smaller but fully phonemically segmented IFA-corpus [25].…”

“…All compiled data tables are fed into a PostgresSQL database for subsequent data mining. Some initial results about speaking rates and phoneme durations are already available [25], but more intricate questions can hopefully be solved using the powerful query language SQL. The corpus is freely available and accessible on line under the GNU General Public License (http://www.fon.hum.uva.nl/ IFAcorpus/).…”