THISL: Thematic indexing of spoken language

• Forschungsstelle: EU FRP
• Projektnummer: 96.0358
• Projekttitel Englisch: THISL: Thematic indexing of spoken language

Schlüsselwörter (Englisch)

Speech recognition; semantic indexing; automatic indexing and retrieval news broadcasts

Alternative Projektnummern (Englisch)

EU project number: EP 23.495

Forschungsprogramme (Englisch)

EU-programme: 4. Frame Research Programme - 1.3 Telematic systems

Kurzbeschreibung (Englisch)

See abstract

Weitere Hinweise und Angaben (Englisch)

Full name of research-institution/enterprise:
Institut dalle Molle d'intelligence artificielle perceptive IDIAP
Research Institute

Partner und Internationale Organisationen (Englisch)

Sheffield University (UK), BBC (UK), Faculté Polytechnique de Mons (B), SoftSound (UK), Thomson-CSF (F), and Intl. Computer Science Institute (ICSI, Berkeley, Subcontractor)

Abstract (Englisch)

The overall objective of the THISL project was to produce a demonstrator system for content-based indexing and retrieval of TV and news broadcasts. This has been achieved: at the project end a demonstrator containing 1800 hours of automatically transcribed and indexed BBC news output (updated daily) had been installed and made available on the BBC intranet. Evaluation both by users within the BBC and through an international evaluation programme conducted by the US National Institute of Standards and Technology (NIST) have indicated that the THISL system is at the state-of4he-art technologically, and offers a valuable new service to archive users, with the promise of a much broader domain of applicability.

The THISL system is based on a connectionist/statistical speech recogniser for radio and TV broadcasts, which produces a word-level transcription. Indexing and probabilistic retrieval techniques are then used, to enable users to search for new items of interest to them. The interface of the demonstrator is similar to that of a web search engine; an alternative spoken query interface was also developed.

In case of space limitation the following can be

The principal technical achievements of the projects are as follows:

1. Development of a broadcast news retrieval demonstrator, installed and in daily use at the BBC.

2. Development of an alternative spoken query interface, using natural language processing technology to recover from recognition errors.

3. Development of a speech recognition system for British English broadcast speech.

4. Investigation of novel information retrieval strategies based on latent semantic
analysis: this was one of the 'flail' contributions of IDIAP.

5. Development and implementation of algorithms for tracking speakers in radio and TV broadcasts.

6. Development and implementation of a novel decoding algorithm for large vocabulary speech recognition to enable real-time recognition of broadcast speech with a relatively low memory overhead.

7. Development and implementation of query expansion and segmentation algorithms for information retrieval.

8. Development and implementation of new probabilistic confidence measures on speech recogniser output: contributions from IDIAP.

9. Adaptation of the basic THISL demonstrator to French language broadcast news:

contributions from IDIAP.

10. Investigation of novel approaches to speech/music discrimination.

11. Development of a keyword spotting algorithm based on a new technique referred to as Iterative Viterbi Decoding: developed by IDIAP.

Each of these areas resulted in (a) software modules that were available for incorporation in the THISL system, and (b) scientific papers published in leading journals and conferences.

Datenbankreferenzen (Englisch)

Swiss Database: Euro-DB of the
State Secretariat for Education and Research
Hallwylstrasse 4
CH-3003 Berne, Switzerland
Tel. [phone]
Swiss Project-Number: 96.0358