A Natural-Language Speech Interface Constructed Entirely as a Set of Executable Specifications

AAAI 1999

TL;DR

This paper introduces SpeechNet, a system that makes the vast information on the web accessible via speech. It achieves this by creating "sihlos" – speech-accessible, hyperlinked objects, each with its own small, highly accurate grammar. The core innovation is constructing these speech interfaces entirely as executable specifications, simplifying development and improving recognition accuracy.

What Problem Does This Paper Address?

The internet, a massive repository of knowledge, is predominantly visual. This creates a significant barrier for visually-challenged users. While speech interfaces exist, building accurate ones for large, complex knowledge bases is challenging. This difficulty stems from the dual task of designing a robust grammar and developing a reliable speech recognizer that can handle diverse inputs. Existing approaches often require extensive user training and struggle with accuracy.

Methods

The paper proposes a novel approach to building speech interfaces by treating them as executable specifications. This means the grammar and the language processor are defined as code that can be directly executed. The system, called SpeechNet, uses "sihlos" (speech-accessible hyperlinked objects) deployed over the internet. When a speech browser accesses a sihlo, it downloads a specific grammar file, allowing the browser to configure itself for high recognition accuracy.

The key technique is the use of an attribute grammar defined in a language called W/AGE (Windsor Attribute Grammar Programming Environment). This specification includes:

  • A dictionary mapping words to database constructs or other phrases.
  • Production rules (syntax) with semantic actions to compute attribute values.
  • Chomsky-like syntactic re-write rules.
  • Equations defining the semantic model.

This "attribute-grammar programming paradigm" allows the language processor and speech recognizer to be built as a single, executable entity.

Datasets

Not discussed in the paper. The "solar man" sihlo is described as an interface to a database containing information about planets, moons, and their discoverers, but the specific dataset used is not detailed.

Key Insights and Ideas

  • Executable Specifications for Speech Interfaces: Building speech interfaces as executable specifications simplifies development and directly links syntax and semantics, leading to more accurate recognition.
  • Modular Grammars (Sihlos): Dividing knowledge into small, hyperlinked modules (sihlos), each with a small, specialized grammar, enables highly accurate, user-independent continuous speech recognition.
  • Web-like Navigation for Speech: The hyperlinking of sihlos allows users to navigate complex knowledge bases using spoken prompts, mirroring the web's visual navigation.
  • Democratizing Speech Interface Development: By providing a framework for executable specifications, the authors aim to make building sophisticated speech interfaces more accessible.

Key Results

The paper highlights the success of the "solar man" sihlo, which can answer thousands of spoken, pseudo-natural-language questions about the solar system. This sihlo was constructed as an 800-line executable specification. The prototype speech browser, written in Java and using IBM's Via Voice engine, achieved very high recognition accuracy for the solar man example. The response time was real-time, and the downloading/reconfiguration of grammars was imperceptible.

Limitations and Future Work

The authors acknowledge that the full potential of SpeechNet will only be realized with tools to assist in designing grammars, constructing language processors, and deploying sihlos. Current work is focused on using the Common Object Request Broker Architecture (CORBA) for deploying sihlos, which will enhance resource sharing, openness, and scalability. Future work aims to enable sihlos to have "sessions" with users, supporting more natural human-computer speech dialogue. The ultimate goal is to create a vast, hyperlinked body of knowledge accessible via speech interfaces anywhere with internet access.

Summary updated

BibTeX
@inproceedings{1999natural,
  title = {{A Natural-Language Speech Interface Constructed Entirely as a Set of Executable Specifications}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1999},
}
Figure