This paper describes a scalable architecture for real-time speech recognizers based on word hidden Markov models (HMMs) that provide high recognition accuracy for word recognition tasks. However, the size of their recognition vocabulary is small because its extremely high computational costs cause long processing times. To achieve high-speed operations, we developed a VLSI system that has a scalable architecture. The architecture effectively uses parallel computations on the word HMM structure. It can reduce processing time and/or extend the word vocabulary. To explore the practicality of our architecture, we designed and evaluated a complete system recognizer, including speech analysis and noise robustness parts, on a 0.18-/spl mu/m CMOS standard cell library and field-programmable gate array. In the CMOS standard-cell implementation, the total processing time is 56.9 /spl mu/s/word at an operating frequency of 80 MHz in a single system. The recognizer gives a real-time response using an 800-word vocabulary.

Scalable architecture for word HMM-based speech recognition and VLSI implementation in complete system

A review of new directions in the evaluation of voice input/output systems is presented. Furthermore, recent studies on human factor aspects involved in speech communication are examined, with the idea that human factors research can improve current voice input/output devices and help to experimentally determine implementation guidelines for developing voice interface design. >

New directions in the evaluation of voice input/output systems

This thesis deals with modelling data existing in large speech corpora using an object-oriented paradigm which captures important linguistic structures. Information from corpora is transformed into objects and are assigned properties regarding their behaviour. These objects, called speech units, are placed onto a multi-dimensional framework and have their relationships to other units explicitly defined through the use of links. Frameworks that model temporal utterances or atemporal information like speaker characteristics and recording conditions can be searched efficiently for contextual matches. Speech units that match desired contexts are the result of successful linguistically motivated queries and can be used in further speech processing tasks in the same computational environment. This allows for empirical studies of speech and its relation to linguistic structures to be carried out, and for the training and testing of applications like speech recognition and synthesis. Information residing in typical speech corpora is discussed first, followed by an overview of objectorientation which sets the tone for this thesis. Then the representation framework is introduced which is generated by a compiler and linker that rely on a set of domain-specific resources that transform corpus data into speech units. Operations on this framework are then presented along with a comparison between a relational and object-oriented model of identical speech data. The models described in this work are directly applicable to existing large speech corpora, and the methods developed here are tested against relational database methods. The object-oriented methods outperform the relational methods for typical linguistically relevant queries by about three orders of magnitude as measured by database search times. This improvement in simplicity of representation and search speed is crucial for the utilisation of large multi-lingual corpora in basic research on the detailed properties of speech, especially in relation to contextual variation.

Object-based modelling for representing and processing speech corpora

The CTH speech database: An integrated multilevel approach

This paper describes the approach taken at Chalmers University of Technology in building up an integrated multilevel speech database for the purpose of speech research and the development of speech coding techniques. The material comprises today isolated speech sounds (phones and diphones) as well as short, semantically unrelated sentences and coherent texts. Data collection is, to start with, restricted to Swedish material and read speech. Registration of the speech samples was carried out under optimal conditions (sound-insulated, anechoic studio) using digital recording equipment (SONY PCM-F1). Segmentation, classification and labeling is performed at eight interlacing levels of linguistic (including acoustic, phonetic and prosodic) analysis.

The CTH-Speech Database: An Integrated Multilevel Approach

A database of spoken Japanese sound has been collected for use in designing and evaluating algorithms for automatic speech recognition. The database is composed of 323 words. It is a special feature of this database that all samples are uttered four times by each speaker (i. e. four tokens per word). Seventy-five male and 75 female data are collected at 15 recording places. Speaker data include sex, age, height, etc. Fifteen research institutions and private enterprises engaged in speech research and development have taken part in the data collection. This is a result of four years' effort by a committee supported by JEIDA (Japan Electronic Industry Development Association). Part of the database has been distributed among the members of the committee.

S. Itahashi

Papers

A Japanese language speech database