Implementation Report -- Speech Recognition Grammar Specification Version 1.0

Lucent Speech Solutions has implemented the W3C Speech Recognition Grammar Specification in full. The specification is well written and rich with examples and details. Lucent has embedded a GRXML/ABNF grammar compiler in its VoiceXML 2.0 compliant Speech Server product. The Lucent Speech Server is designed for high density scalable solutions, where each server is capable of handling up to 192 simultaneous ASR and TTS channels.

All six user agents are implemented, as listed below:

• XML Grammar Processor for ASR (user agent)
• XML Grammar Processor for DTMF (user agent)
• ABNF Grammar Processor for ASR (user agent)
• ABNF Grammar Processor for DTMF (user agent)
• XML text parser (command-line user agent)
• ABNF text parser (command-line user agent)

In addition, an XML -> ABNF converter (grammar processor) has been implemented.

Note 1: Unless noted otherwise, the root rule is imported for all grammar documents.

Note 2: Quotes were added around any token that contained whitespace in the meta in.X declarations to simplify tokenization in the automated testing procedure. In abnf-keywords.gram, for example,

    meta 'in.1' is 'lexicon and public base';

was replaced by

    meta 'in.1' is 'lexicon and "public base"';

Note 3: Although the grammar processors can handle any combination of languages, our ASR engine can currently only provide missing pronunciations in the default language. If a grammar with en as the default language contained both "xyz!en" and "xyz!es", for example, "xyz!es" would have to be provided in a lexicon while "xyz!en" could be pronounced using letter-to-sound rules.

Microsoft has implemented the XML Form of the Speech Recognition Grammar Specification (SRGS) in a developmental version of the Microsoft Speech API (SAPI).

SAPI is an API which allows developers to use speech recognition and/or speech synthesis in their applications. To conduct the tests specified in the implementation report, a text-input interface was used with SAPI's core grammar processor. Given SAPI's ability to use the core grammar processor in speech, DTMF and text environments, we believe this successfully demonstrates coverage of SRGS in the following agents required by the Implementation Report Plan :

• XML Grammar Processor for ASR
• XML Grammar Processor for DTMF
• XML text parser

Previous versions of SAPI may be found at http://www.microsoft.com/speech/.

As a founding member of the Speech Application Language Tags (SALT) Forum (http://www.saltforum.org) and an active member of the W3C Voice Browser and Multimodal Interaction Working Groups, Microsoft believes speech standards will play a key role in the growing market for speech applications. We consider SRGS a thorough, well-designed specification which, by providing a common syntax for speech recognition grammars, should help promote interoperability and portability in speech application development. Microsoft intends to support SRGS in its suite of forthcoming SALT products:

• .Net Speech SDK, a set of speech application development tools and speech controls integrated with Visual Studio.NET that enables rapid incorporation of speech into web applications (a Beta 2 version of the SDK is available at http://www.microsoft.com/speech/);
• .Net Speech platform, an integrated multimodal and telephony platform for multiple clients such as PCs, telephones, wireless personal digital assistants (PDAs), and Tablet PCs.

As required by the Implementation Report plan, some technical details of the SAPI implementation are as follows:

1. Relationship of SAPI output to LPS
   SAPI's XML representation of recognition output was mapped explicitly to the LPS defined in Appendix H by a recursive traversal of the parse tree. In some tests a complete mapping into LPS was not always possible, for example, the content of the tag element and the exact path of external rules are not copied directly. However, since these are minor aspects of only the surface form of LPS (itself an informative part of the specification) and they in no way affect the behaviour of the grammar processor as defined in the specification, we do not consider these an unsuccessful implementation of the tests.
2. Weights and probabilities
   As noted in the Implementation Report Plan, pass/fail testing of weights on alternatives and probabilities on repeats is not possible. We believe these features are implementable and useful. We have implemented support for both weights and repeat probabilities into our ASR grammar processor. We believe that, when properly estimated, weights and repeat probabilities have a positive effect in maximizing recognizer performance.
3. Language support
   Speech recognition engines were not available for certain languages required by the tests. For those tests where a single language is used in the grammar and a recognition engine was not available for that language, SAPI's text input mechanism was used with the grammar processor to compile the grammar, parse the input and produce successful output, and we have considered this a successful implementation.
4. Test set version
   The SAPI implementation was run on the set of tests posted at: http://www.w3.org/2002/06/srgs-irp/ on 24 October 2002 (see http://lists.w3.org/Archives/Public/www-voice/2002OctDec/0039.html for details).
5. Unimplemented features
   As noted in the results of the tests, the following features of the specification have not been implemented.
   • multiple languages within the same grammar
     This appears to be a useful feature for certain deployment scenarios.
   • lexicon content
     Although we have implemented the syntax of <lexicon> (and thereby the correct test set behaviour), we have not implemented the semantics of lexicon look-up. The ability to specify pronunciations is clearly a very useful feature.
   • repetition of tag elements equivalent to a single tag element (tag-repetition.grxml.)
     In our belief, where tag elements are repeated without tokens for input, this is a developer error, and platforms should not be required to resolve the repetition to a single tag.

IBM has implemented the W3C Speech Recognition Grammar Specification, both XML Form and ABNF in its WebSphere Voice Server VoiceXML browser products, as well as in its Java Speech API implementation. The results are attached.

The following features are not supported :

• Special Rule[2.2.3]: $GARBAGE
• Reference to external N-Gram documents [2.2.4]
• Language references [2.2.5]
• Weights for alternatives [2.4.1]
• Repeat Probabilities [2.5.1]
• Mixed Language [2.7]
• Tag format[4.8]
• Reference to external Lexicon documents [4.10]
• Meta and HTTP-Equiv[4.11.1]

BeVocal has implemented the XML and ABNF form of the Speech Recognition Grammar Specification [SRGS] in the Grammar module of its VoiceXML 2.0 Interpreter. The VoiceXML 2.0 Interpreter is the core of the BeVocal Foundation Platform which is currently hosting Carrier and Enterprise voice applications.

BeVocal has provided advanced VoiceXML 2.0 and grammar (XML and ABNF) features to our developer community through frequent releases to BeVocal Cafe (http://cafe.bevocal.com). This has enabled BeVocal to provide real implementation feedback about SRGS to the W3C Voice Browser Working Group.

The SRGS Candidate Recommendation is a significant milestone to the voice industry and also the voice application development community. Not only is it already implemented and used in production by a number of voice solutions providers like BeVocal, but it is well-written, concise, with clear examples for developers.

In addition to an XML form, SRGS provides a more concise ABNF form for grammar developers. At BeVocal, both forms create the same internal representation, which is a testament to how well the two forms are aligned in design and features.

BeVocal is providing the following Implementation Report on a specific set of ABNF features which have been implemented and tested on an internal prototype. An estimated production release for these features is in the next release following our Q1 release.

Note also that BeVocal has implemented rule weights and repeat probabilities as described in SRGS, in both our ABNF and XML implementation.

SpeechWorks has implemented the W3C Speech Recognition Grammar Specification fully in OpenSpeech Recognizer, our high performance network-based speech recognition engine that is optimized for VoiceXML applications. Through this native implementation we provide an easy path for telephony platform vendors to support the grammar specification in their own DTMF- and ASR-capable products. The SpeechWorks implementation features complete multilingual capabilities allowing arbitrary grammatical constructs using an unlimited number of languages and Unicode character encodings, along with embedded ECMA-262 scripts for application-specific search pruning and semantic result computation. Based on our own experience and feedback from many partners and customers, we give a strong endorsement to SRGS as practical, expressive, and complete.

Note: While we did implement language attachments on rule references, this feature was found to be unnecessary and confusing to many users. We support its removal.

Loquendo believes in the W3C standardization activity for the global speech technology market and announces the W3C Speech Recognition Grammar Specification implementation in Loquendo ASR product, which is part of the high-quality Loquendo Speech Suite. Loquendo's Speech Suite offers all that is necessary for developing any kind of flexible speech-enabled services. It enables you to choose from Loquendo's technological components range including Loquendo TTS, Loquendo ASR and VoiceXML Interpreter, and exploits them using a flexible integration infrastructure.

The current implementation fully supports the XML format of W3C SRGS (ABNF format is not yet supported) and it is being extended to support DTMF input.

Loquendo implementation was run on the official test set ("srgs-test-set-20021017.zip") by applying the following changes: ignore the removed tests ("tag-repetition", "lang-ruleref") and manual update to the latest tests regarding media-type handling. The tests ran on a development version of the grammar processor which allows a better internal LPS output structure.

Technical details of Loquendo implementation are the following:

• Weights and probabilities were fully implemented and tested and they will help to improve large scale voice application.
• Multiple language capability is certainly a very useful feature for world-wide applications. E.g. Loquendo exploited this feature on a bilingual Spanish Castillian and Catalan application.
• Korean and Chinese tests has been marked "no-lang", as they are not yet supported by Loquendo ASR engine. However, during offline testing they were successfully compiled and the expected output was produced.
• Although the implementation passed the test on Base-URI feature ("base-declaration"), Loquendo has some concerns on using in tests a relative URI on "xml:base" attribute. See Section 5.2 of RFC-2396.
• The lexicon element is not currently supported in the SRGS implementation; waiting the standard phonetic lexicon specification. This functionality is allowed at a platform service level.