Skip to main content

Adverse-Environment Recognition of Speech

Objective

The objective of ARS was to develop improved algorithms for medium-size vocabulary speaker-dependent speech recognition in the presence of noise, and to build a real-time demonstrator. The demonstrator was to incorporate an isolated word noise-robust recogniser, verify algorithm performance, and address the problem of speech-based person-machine dialogue as a system interface in practical applications. The application environment chosen was the car.
The aim of the project is to extend the state of the art in speech recognition and to place this innovative technology in adverse environments such as car and factory floor. Starting from an established base of expertise, this project involves theoretical work on algorithms and the development of hardware prototypes. To get the best recognition performance, algorithms covering the different aspects of signal processing were considered. The activities were subdivided into 6 work packages concerning respectively system definition and standards, transducers and noise reduction, feature extraction, pattern processing, human factors and user interface, system prototyping and evaluation. After a brief presentation of the general structure of the project (objectives, organisation, participation, resources), this paper presents the state of the work after two years.

The objective of adverse environment recognition of speech (ARS) project was to develop improved algorithms for speech recognition in the presence of noise and to build a real time demonstrator. The demonstrator was to incorporate an isolated word noise robust recognizer, verify algorithm performance, and address the problem of speech based person machine dialogue as a system interface in practical applications.

The application environment chosen was the car. The system has a 100 word vocabulary, chosen by each national group of partners and tailored to the specific application environment. Advances were made in:
reduction, by signal preprocessing, of the effects of noise on speech signals;
feature extraction, to improve noise robustness;
study and refinement of algorithms for speech pattern matching in noisy environments;
speaker adaptation;
dynamic system adjustment to user feedback and the development of error correction strategies in the human interface;
development of system prototypes (hardware and firmware) for real time speech recognition.

The real time demonstrator was based on a general purpose digital signal processing (DSP) chip attached to a personal computer or a stand alone system. A multilingual database collected in noisy environments was made available and used for the evaluation of baseline systems. These were realized according to a common standard suitable for exchanging the software modules of the algorithms. Various algorithms were developed and evaluated and a set of algorithms for the final prototype were selected. A human machine interface concept was defined and the porting of the various models to the target system hardware was initiated.
The complete chain of processing has been initiated on a real time hardware; 2 demonstrators have been installed inside cars for assessment of their performance in real operating conditions.
The requirements included a 100-word vocabulary, chosen for each language group of partners and tailored to the specific application environment. Advances were needed in terms of:

- reduction, by signal preprocessing, of the effects of noise on speech signals
- feature extraction, to improve noise robustness
- study and refinement of algorithms for speech pattern matching in noisy environments
- speaker adaptation
- dynamic system adjustment to user feedback and the development of error correction strategies in the human interface
- development of system prototypes (hardware and firmware) for real-time speech recognition.

The system would be integrated in a real-time demonstrator based on a general-purpose DSP chip attached to a personal computer on a stand-alone system. Performance evaluations were first scheduled in the laboratory, using databases collected in noisy environments, to evaluate the resulting rate of correct recognition. Performance under field conditions were then to be assessed from a prototype fitted in a car and a laboratory system installed in a factory.

Coordinator

Centro Studi e Laboratori Telecomunicazioni SpA
Address
Via G. Reiss Romoli, 274
10148 Torino
Italy

Participants (9)

Logica Ltd
United Kingdom
Address
64-68 Newman Street
W1A 4SE London
Logica Ltd
United Kingdom
Address
Cobham Park Downside Road
KT11 3LX Cobham
Matra Communication
France
Address
Rue Jean-pierre Timbaud
78392 Bois D'arcy
PAGE IBERICA
Spain
Address
Lagasca 88
28001 Madrid
POLITECNICO DI TORINO
Italy
Address
Corso Duca Degli Abruzzi 24
10129 Torino
TELECOM PARIS
France
Address
46 Rue Barrault
75634 Paris
UNIVERSITAT POLITECNICA DE MADRID
Spain
Address
Campus De Montegancedo
28660 Madrid
University of Cambridge
United Kingdom
Address
Trumpington Street
CB2 1PZ Cambridge
University of Keele
United Kingdom
Address

ST5 5BG Keele