Project Name | Start Date | End Date |
---|
Bayesian Music Transcription | 1997-09-01 | 2004-09-01 |
- Description
- Muziektranscriptie is het omzetten van een gedigitaliseerde opname van een muziekuitvoering in een beschrijving die te lezen en te begrijpen is. Transcriptie van iedere willekeurige muziekuitvoering zonder enige aanname van de soort muziek is een zeer moeilijk probleem. Het doel van het proefschrift van de heer Cemgil was om een praktische aanpak voor muziektranscriptie te ontwikkelen door een combinatie van kennis vanuit muziekwetenschap en akoestiek, en computationele technieken afkomstig uit de kunstmatige intelligentie.
Music Transcription is the conversion of a digitized recording of a musical performance in a description which is to read and understand. Transcription of any given music performance without any assumption of the kind of music is a very difficult problem. The aim of the thesis of Mr. Cemgil was to develop a practical approach to music transcription by combining knowledge from musicology and acoustics, and computational techniques derived from artificial intelligence.
- Disciplines
- Musicology
Computer Science
- Institutions
Donders Centre for Cognition (Secretariat) Radboud Universiteit Nijmegen Donders Centre for Cognition - Persons
Prof.dr. H.J. Kappen (Co-supervisor) P.J.T.M. Trilsbeek (Researcher) Prof.dr.ir. P.W.M. Desain (Project leader) Prof.dr. H.J. Honing (Project leader) Prof.dr. C.C.A.M. Gielen (Supervisor) Dr. A.T. Cemgil (Doctoral/PhD student)
|
Prosodic realizations of text structure | 1998-02-01 | 2004-12-01 |
- Description
- Just as a writer brings structure to his written text using comma's, dots and capitals, a speaker brings in his narration structure using so-called prosodic means: pausing between words and phrases, speech rate, intonation, emphasis and volume. The study of Hanny den Ouden provides insight into the complex process of speaking and understanding between speakers and listeners. This knowledge can also be improved in text-to-speech systems, allowing computers texts can speak in a more natural way.
- Disciplines
- -
- Institutions
Department of Culture Studies Tilburg (Secretariat) - Persons
Prof.dr. L.G.M. Noordman (Supervisor) Dr. J.M.B. Terken (Supervisor) Dr. J.N. den Ouden (Doctoral/PhD student)
|
Computational Comparison en Classification of Dialects | 1998-08-01 | 2004-01-01 |
- Description
- Doel is het meten en classificeren van taalvariatie. Eerst meten we de fonetische afstanden tussen dialecten. Vervolgens classificeren we de dialecten op basis van hun onderlinge afstanden. In eerste aanzet wordt gebruik gemaakt van data uit de Reeks Nederlandse Dialectatlassen. Vergelijking vindt plaats met het Levenshtein-algoritme, dat de kosten bepaalt van het veranderen van het ene woord in het andere door toevoeging, verwijdering en vervanging van klanken.
"This project researches several methods for comparison of dialects. These methods delivers distances between dialects. On the basis of these distances the dialects can be classified. Therefore, at the same time several methods for classification are researched. This project will try to give an answer to the following dialectological question: in which way do we get the most reliable dialect classification? In 1988 Hoppenbrouwers presented his feature frequency method. In 1995 Kessler used Levenshtein distance for comparing Irish dialects. Later, Levenshtein distance was used for Dutch dialects by Nerbonne et al. (1996) and Nerbonne and Heeringa (1997). The results were promising. An intermediate form is the word per frequency method. Per word, the frequencies of phones or features of phones are counted. Dialects are compared by comparing their frequencies. In this research, the Levensthein distance will be refined and elaborated. As in previous research, the dialects will be classified by clustering (Jain and Dubes, 1988), multidimensional scaling (Kruskal and Wish, 1978) and Indscale (Kruskal and Wish, 1978). With these methods results are readily obtained. To get an indication of the reliability, results are compared with traditional and experimental dialectology. For comparison with the results of the traditional dialectology several dialect maps will be used. For Dutch dialects, the maps of Te Winkel, van Ginneken, Weijnen, Daan (Daan and Blok, 1969) are employed. If available, English or American dialect maps will also be used. For comparison with experimental dialectology, a recording of a number of dialect fragments is constructed. Test persons listen to the tape, where each test person expresses the difference with respect to a standard language in a number (Van Hout en Münstermann, 1981, Gooskens, 1997). Now, the correlation between computational distances and experimental distances can be calculated."
- Disciplines
- Language studies
- Institutions
Department of Humanities Computing RUG (Secretariat) Rijks Universiteit Groningen - Persons
Dr. R.A.M.G. van Bezooijen (Researcher) Dr. T. de Graaf (Researcher) Prof.dr.ir. J. Nerbonne (Supervisor) Dr. W.J. Heeringa (Doctoral/PhD student)
|
MIMORE | 2000 | 2016 |
- Description
- The MIMORE tool enables researchers to investigate morphosyntactic variation in the Dutch dialects by searching three related databases with a common on-line search engine. The three databases involved are DynaSAND, DiDDD and GTRP.
- Disciplines
- Computational Linguistics
- Institutions
Meertens Institute Research and documentation of Dutch language and culture (Collaboration) Royal Netherlands Academy of Arts and Sciences - KNAW - Persons
Meertens (Project leader)
|
Broadcast Restoration of Archives through Video Analysis | 2000-00-01 | 2004-09-01 |
- Description
- The Brava project (Broadcast Restoration of Archives through Video Analysis) is an European project that aims at developping tools for digital restoration of large amounts of broadcast archive documents (video and film), for re-exploitation purposes. The project is of considerable importance for the preservation of the European cultural heritage. It will provide tools fore re-exploiting the huge archives programme stocks available from archives such as Institut National de l'Audiovisuel (Paris), Britisch Broadcast Corporation (UK), Portuguese Radio-Television (Portugal). The project will focus on high-level video analysis techniques and will build upon earlier developments realised within the framework of the previous Aurora project.
- Disciplines
- Cultural Heritage
Media Studies
- Institutions
Advanced School for Computing and Imaging - ASCI (Secretariat) University of Delft - Persons
Prof.dr.ir. J. Biemond (Supervisor) MSc A. Rares (Doctoral/PhD student)
|
Integrating linguistic and statistical information for spoken language processing | 2000-01-01 | 2004-01-01 |
- Description
- The project aims at bringing together the knowledge-based approach to linguistic research and the statistical approach to speech research for the development of language models for speech recognition. The role of a language model in speech recognition is to determine what can be said and with what probability. Current language models of this sort are based on a very crude linguistic model, the so-called n-gram model. This model simply conditions the probability of each word on a small number of preceding words. Although there have been many attempts to improve language models by incorporating linguistic knowledge into the n-gram model, these methods have not achieved a significant improvement. There is a growing consensus that the n-gram model is reaching the limit of its possibilities and that there is a need for searching for new directions.
- Disciplines
- -
- Institutions
Department of History - UvA Faculty of Arts and Social Sciences (Secretariat) - Persons
Prof.dr.ir. R.J.H. Scha (Project leader) Prof.dr. L.W.M. Bod (Project leader)
|
Computational Models of Multimodal Dialogues | 2000-01-01 | 2003-09-01 |
- Description
- In this project we study models of dialogue which are formalised and well defined, and thus computationally tractable and potentially of practical use. The focus will be on dialogues where spoken language is one of the in- or output modalities. In 2000 we hope to address the following issues:
(1) Memory-based spotting of communication errors. One of the major problems with the current generation of spoken dialogue systems is that they are insufficiently able to monitor the status of the ongoing conversation. In the past year, we (Krahmer, Swerts, Theune and Weegels) have worked on error-spotting in human-machine communication. This work has shown that humans signal errors using a number of cues (e.g., marked intonation, longer utterances, disconfirmations, repetitions, marked word-order, etc.). A dialogue system which is able to spot such cues automatically might be better able to signal and respond to errors. We would like to investigate to what extent the various cues can be detected automatically from a word-graph (the output of a speech recognition engine).
(2) Changing modalities to solve errors. Of course spotting errors is one thing, solving them is quite another. It seems that one promising way to solve errors is by changing modalities. For instance, presenting visual feedback and allowing for users to select an object via a different modality than speech (e.g., touch) may help in solving problems.
(3) Generating descriptions in multi-modal contexts. Beun and Cremers have investigated the way humans refer to objects in multi-modal contexts. It seems that the algorithms developed by Krahmer and Theune for the generation of descriptions are very well-suited to model the findings of Beun and Cremers regarding human behaviour.
- Disciplines
- Linguistics
Speech technology
- Institutions
Universiteit Tilburg - Persons
Prof.dr. E.J. Krahmer (Project leader)
|
Automatic speech recognition for Dutch dysarthric speech: a pilot study | 2000-01-01 | 2003-12-01 |
- Description
- The goal of this pilot study was to conduct a feasibility study into automatic speech recognition (ASR) of Dutch dysarthric speech. To find out to what extent ASR can be used to the benefit of dysarthric speakers, we intended to answer the following questions: * How well can dysarthric speech be recognised by a continuous speech recogniser (CSR) trained on non-dysarthric speech? * Will the recognition results improve if we train the CSR on (a limited amount of) speech of dysarthric speakers? * To what level of complexity are automatic recognition tasks of dysarthric speech feasible with current ASR technology? - Methodology: We conducted a series of experiments for which we used read speech of two dysarthric speakers and two non-dysarthric (reference) speakers. All speakers read numbers, frequent utterances, semantically unpredictable sentences and phonetically rich sentences. For recognition a standard phone-based CSR was used. Results of the recognition experiments were presented as word error rates, i.e. the number of substitutions, insertions and deletions dived by the total number of words. Performance of the CSR was tested by using speaker independent models, speaker dependent models (using a jack-knife procedure) and by extending the size of the lexicon and the amount of training material. - Results: Word error rates ranged from 4.5% to 41.0% for dysarthric speech that is recognised with the CSR trained on non-dysarthric speech. Relative large improvements were found (50% to 100%) when the CSR was trained on speech of the dysarthric speakers. Results indicate that ASR of dysarthric speech is possible for low-perplexity tasks, i.e. when using a language model. ASR of dysarthric speech also seems promising for higher perplexity tasks, especially when speech rate of the speakers is relatively slow.
- Disciplines
- Speech technology
- Institutions
St. Martin Clinic (Secretariat) Radboud Universiteit Nijmegen - Persons
Prof.dr. L.W.J. Boves (Researcher) Prof.dr. A.C.M. Rietveld (Researcher) Prof.dr. A.C.H. Geurts (Researcher) P. Holtus (Researcher) Dr. J. van Limbeek (Researcher) Drs. E. Sanders (Researcher) M.B. Ruiter (Researcher) Dr. L.J. Beijer (Researcher) Dr. W.A.J. Strik (Project leader)
|
Linguistic search tool for bare text corpora | 2000-02-01 | 2005-02-01 |
- Description
- In dit onderzoeksproject wordt gewerkt aan de specificatie en implementatie van een tool-kit voor reguliere expressies en finite-state automaten. De tool-kit wordt vervolgens gebruikt om aan de hand van onderzoeksprojecten te komen tot de constructie van een systeem voor het zoeken naar taalkundig interessante patronen.
- Disciplines
- -
- Institutions
Department of Humanities Computing RUG (Secretariat) - Persons
Prof.dr. G.J.M. van Noord (Project leader)
|
Linguistic knowledge and word sense disambiguation | 2000-04-01 | 2004-10-01 |
- Description
- In dit onderzoek staat de vraag centraal hoe bij de computationele verwerking van natuurlijke taal meerduidigheid moet worden behandeld. Hiertoe worden een aantal bestaande modellen als uitgangspunt genomen en toegepast op (a)lexicale afhankelijkheidsstructuren en (b) de Nederlandse taal.
- Disciplines
- Linguistics
- Institutions
Centre for Language and Cognition Groningen - CLCG (Secretariat) - Persons
Prof.dr. G.J.M. van Noord (Project leader) Prof.dr.ir. J. Nerbonne (Supervisor) Dr. T. Gaustad (Doctoral/PhD student)
|