The IFA Spoken Language corpus is a free (GPL) database of hand-segmented Dutch speech. It 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 (41 minutes/speaker), 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.
The IFA Spoken Language corpus is currently in version 1.0. This is the "reference" version and the first I consider consistent enough to be usefull. However, the annotations (labeling) still contains errors. This means that there are inconsistencies in a few percent of the labels (e.g., wrong word assignment of syllables/phonemes, stress errors, etc.).
| Sex | Age | ID | Recorded sentences | Segmented sentences |
| F | 20 | N | 3736 | 2760 |
| F | 28 | G | 4180 | 3978 |
| F | 40 | L | 3112 | 2485 |
| F | 60 | E | 4181 | 3245 |
| M | 15 | R | 2125 | 1439 |
| M | 40 | K | 2720 | 1891 |
| M | 56 | H | 2894 | 2368 |
| M | 66 | O | 3781 | 1696 |
| Total | - | - | 26733 | 19867 |
A more extended paper describing the corpus can be found here (pdf) and here (pdf). This has been presented as a poster at the EUROSPEECH2001 conference.
The IFA spoken language corpus is constructed using the Praat speech editting and analysis program. All speech material is accessible with praat.
The Dutch Language Organization (Nederlandse Taalunie) holds all copyrights (unless explicitely indicated otherwise) and makes the complete corpus available under the GNU General Public License (see below).
The corpus is quite large (>1.8GB), so it might be a good idea to only checkout if you want all. If not, try wget instead:
The copyrights to all materials presented here are owned by the
Nederlandse Taalunie
and R.J.J.H. van Son unless explicitely stated otherwise.
All materials are licensed under the
GNU General Public License (GPL).
For information or further details on
licensing see here or contact me
at the email address above. A small history file is available.
The corpus contains speech from 8 Dutch speakers (others are still not processed). For each speaker, a Fixed text has been recorded in several "styles", and a retold version of the fixed text. Furthermore, each speaker told an Informal story face-to-face with an interviewer which was the basis of a speaker specific variable text corpus, which was read and retold by each speaker individualy.
| Sex | Age | ID | Recorded sentences | Segmented sentences |
| F | 20 | N | 3736 | 2760 |
| F | 28 | G | 4180 | 3978 |
| F | 40 | L | 3112 | 2485 |
| F | 60 | E | 4181 | 3245 |
| M | 15 | R | 2125 | 1439 |
| M | 40 | K | 2720 | 1891 |
| M | 56 | H | 2894 | 2368 |
| M | 66 | O | 3781 | 1696 |
| Total | - | - | 26733 | 19867 |
You can find information on the label protocol, the naming conventions, and the phoneme labels here (in Dutch) and here (some of it in English).
Speaking styles are:
Recordings were made on two separate channels, from a fixed microphone (_fm, Sennheiser MKH 105, a HF condenser microphone) and from a head mounted microphone (_hm, Shure SM10A, a moving coil microphone). Next to the sound files, some analysis results are available, e.g., the spectral Center of Gravity (CoG) files, the Intensity curve, and the Pitch. The Spectral Center of Gravity (= the first spectral moment) is a compact way to represent spectral changes, especially changes in speech source (noise and voicing) and changes related to articulatory movements. These are stored as Praat ASCII Sound files. Sound files are CD quality (directly recorded on Audio CD), i.e., 44.1 kHz and 16 bit.