Loquendo ASR

Automatic Speech Recognition

Loquendo ASR is the next-generation speech recognition technology for speech-enabled applications. It is speaker-independent and reliably recognizes large-scale vocabulary continuous speech, even in the noisiest environments, including wireless. Loquendo is the only speech technology vendor that provides a complete product line for servers, desktop, PDA and embedded, guaranteeing the same wide range of languages and the same core engine in all these environments. Loquendo ASR currently powers services that handle millions of phone calls, such as fully automated directory assistance services, mobile public voice portals, and embedded applications.

Bringing you a world of benefits!

Loquendo ASR gives integrators the freedom to create user-friendly services that can be as complex as they want them to be in terms of vocabulary size and interaction flexibility. Loquendo ASR perfectly fits the requirements of each and every application scenario - however complex - in any language! Here are a few of the reasons why: Broad Vocabulary & Flexible Recognition: recognizes up to 1,000,000 words; supports isolated and continuous speech. Highly Accurate Speech Recognition thanks to: - integration of neural networks and hidden Markov models, - highly accurate phonetic transcribers specialized for each language (also used in Loquendos excellent quality TTS), - detailed acoustic-phonetic units trained on large speech corpora, - technological fine-tuning based on field-data. Extended Standards Support: optimized for VoiceXML applications; MRCP and AURORA compliance; complete grammar standard support, both W3C SRGS and Semantic Interpretation. High Efficiency: low-computational power requirements enable a large number of recognition channels to run contemporarily, with both small and large vocabularies. Rapid Extensibility to new languages: the methodology tuned for our wide range of languages is rapidly extended to any other. Powers Loquendo Speaker Verification technology in all available languages.

Technology that is Simple yet Powerful!

A complete set of simple and powerful features guarantees robust speech technology. They enable: Improved Barge-in capability to guarantee both high reactivity and robustness to noise and to background speech. A flexible rejection mechanism, which identifies any linguistic expressions that are not acceptable within a specific domain. Dialogue-flow management, which is achieved through confidence values provided for all the Nbest hypotheses returned on a sentence-by-sentence & word-by-word basis.

ASR Tools
Loquendo ASR provides users with a learning tool package, including: Phonetic Learning for an automatic discovery of pronunciation variants for vocabulary words and for detecting frequent not foreseen linguistic formulations to improve a speech recognition grammar Acoustic Model Adaptation: to improve recognition performance using audio material recorded in the field (environment, speaker, channel adaptation). A Sophisticated Speech Assistant Toolkit guarantees the rapid and efficient definition of Recognition Objects (ROs) and Recognition Packages, such as Grammar R O s, Language Modelling ROs using statistical N-grams.

VOCAL TECHNOLOGY AND SERVICES

Significant memory requirement reduction: ROs can be both permanent (and therefore shared by all recognition channels) and dynamic (i.e. loaded run-time when required and discarded once they have been used). In unpredictable situations, ROs can be created, stored and deleted on the fly. Loquendo ASR also provides: A re-usable built-in grammar library for each language. Phonetic segmentation, which provides the phonetic representation and related time-stamps for each phoneme within a sentence. This is often a prerequisite, especially in avatar animation.

Loquendo ASR - Technical Specifications

Main Features Speaker Independent Open Vocabulary Noise robustness (e.g. in-car, wireless, etc.) Optimized for Telephonic Speech

Basic technology Configurable Recognition Modalities

Neural Networks + Continuous Density Hidden Markov Models Grammar based Continuous Speech Recognition with Statistical Language Modelling Phonetic Decoding N-Best Decoding Confidence Scores at sentence and word level Tuneable Voice Detection sensitivity Improved Barge-In functionalities Grammar handling and fast grammar compilation on the fly Voice enrolled grammar Natural Language Processing Acoustic model composition Optimized for VoiceXML applications

Key features

Tools

Phonetic Learning tool to face user formulations variability and pronunciation variants (non-native speakers and regional accents) Acoustic Model Adaptation tool to improve recognition performance using audio data recorded in the field (environment, speaker, channel adaptation). Italian, Castilian Spanish, French, German, Brazilian, Portuguese, U.K. and U.S. English, Greek, Chilean, Argentinean, Swedish, Catalan, Mexican, Dutch JSGF (Java Speech Grammar Format) W3C SRGS 1.0 (XML and ABNF Form) + SISR 1.0 Network Embedded CE.NET, Pocket PC 2003, Windows XP Embedded, WindowsXP TabletPC, VxWorks, Linux Loquendo API

Supported languages

Grammar formalisms

OS supported

MS Windows (2000, 2003, XP), Linux Red Hat (7.x, 8.0, 9.0) Loquendo API (C/C++ and .Net) MRCP (for Client Server Architecture) Intel Dialogic Audio Source support DSR/AURORA support

Interfaces

CPU requirements

Memory requirements

15 MB per language shared among channels. Few MB per channel depending on the recognition task (e.g. 5 MB for Connected Digits Recognition, 15 MB for a grammar with 10.000 words).

ROM: 4 MB per language RAM: 4 MB or more depending on grammar complexity

To find out how Loquendos offering can position your company for success, please visit www.loquendo.com
2006 - Loquendo. All rights reserved. Loquendo, and all Loquendo logos are trademarks registered by Loquendo. All other trademarks belong to their respective owners. The information contained in this brochure is subject to modification without notice.

Loquendo - Vocal Technology and Services via Valdellatorre, 4 - 10149 Torino - Italy tel. +39 011 2913111 - fax +39 011 2913199 www.loquendo.com info@loquendo.com

01/2006 - LOQUENDO - AMBR 042 ENG

Connected Digits Recognition: 80 channels on an Intel Pentium IV 3.2 GHz CPU Grammar with 10,000 words: 20 channels on an Intel Pentium IV 3.2 GHz CPU

Intel Xscale400MHz (or equivalent), SH4, X86 (more under development)