Introduction

22.2433GPPRelease 17Speech recognition framework for automated voice servicesStage 1TS

Forecasts show that speech-driven services will play an important role on the 3G market. People want the ability to access information while on the move and the small portable mobile devices that will be used to access this information need improved user interfaces using speech input. At present, however, the complexity of medium and large vocabulary speech recognition systems is beyond the memory and computational resources of such devices. Also associated delay to download speech data files (e.g. grammars, acoustic models, language models, vocabularies etc. …) may be prohibitive. Eventually, it may not always be acceptable for the speech service providers to allow download of these speech data files if they contained confidential information (password (security issue), customer names and address (privacy issue)) or intellectual properties; for example a well crafted speech grammar is often considered by speech service providers as a trade secret.

Server-side processing of the combined speech and DTMF input and speech output can overcome these constraints by taking full advantage of memory and processing power as well as specialized speech engines and data files. However, the distortions introduced by the encoding used to send the audio between the client and the server as well as additional network errors can degrade the performance of the speech engines; therefore also limiting the achievable speech functionalities. A server-side speech service is generally equivalent to a phone call to an automatic service. As for any other telephony service, DTMF is a feature that should always be considered as needed.

This document describes a generic speech recognition framework to distribute the audio sub-system and the speech services by sending encoded speech and meta-information between the client and the server. Instead of using a voice channel as in today’s server-based speech services, an error-protected data channel will be used to transport encoded speech from the client audio sub-system (terminal client) to remote speech engines (on server) for processing (e.g. speech recognition, speaker recognition,). The speech recognition framework will also enable downlink data streaming of voice and recorded audio prompt generated by server to the terminal client audio subsystem. The speech recognition framework may use conventional codecs like AMR or Distributed Speech Recognition (DSR) optimized codecs.

The speech recognition framework will provide users with a high performance distributed speech interface to server-based automatic speech services with communication, information access or transactional purposes.

The types of supported user interfaces include those that are voice only, for example, automatic speech access to information, such as a voice portal described in this section. These typically support combined speech or DTMF input.

In the future, a new range of multi-modal applications is also envisaged incorporating different modes of input (e.g. speech, keyboard, pen) and speech and visual output.