DARPA Speech Understanding Research

From MSc Voice Technology
Revision as of 18:38, 17 September 2023 by Page (talk | contribs) (→‎Future research)
Jump to navigation Jump to search

Introduction

DARPA stands for Defense Advanced Research Projects Agency, a research agency of the United States Department of Defense responsible for the development of emerging technologies for use by the military. As the DARPA's then-director George Heilemeir stated, "Get computers to read Morse code in the presence of other code and noise, get computers to identify/detect key words in a stream of speech, [...] make a real contribution to command and control, and; do a good thing in sonar"[1]. Consequently, the project received funding from the U.S. Department of Defense, particularly the Navy, given its potential military applications.

Given the military context, the project was subject to specific requirements. Notably, it needed to recognize multiple speakers simultaneously and achieve real-time speech recognition with no delays. As a result, the project's research objectives were defined as follows:

  • Accepting connected speech
  • Recognizing speech from multiple cooperative speakers
  • Accepting 1000 words
  • Yielding only <10% semantic errors
  • Achieving real-time understanding

It's important to note that the project aimed for understanding rather than mere recognition of isolated words, focusing on comprehending the context and intent behind spoken language rather than isolated word identification.

Historical Context

While there had been partially successful attempts to understand discrete speech (see, for example, Bell's Audrey), there were virtually no systems capable of comprehending continuous speech at the time, except for Raj Reddy's recognition system, which was primarily used for issuing chess commands[2]. Furthermore, previous methods were limited to vocabularies of no more than 200 words (e.g., IBM's 16-word "Shoebox"). In contrast, DARPA's Speech Understanding Research (SUR) project aimed to achieve speech recognition with a vocabulary of at least one thousand words. As a result, the project's goals significantly surpassed the capabilities of existing state-of-the-art solutions.

Project Progess

Research groups were established at Carnegie Mellon University (CMU), SRI International, MIT's Lincoln Laboratory, Systems Development Corporation (SDC), and Bolt, Beranek, and Newman (BNN). CMU's research efforts resulted in the development of two systems, HARPY and HEARSAY-II, while BNN was responsible for creating Hear What I Mean (HWIM).

HARPY

Harpy demonstrated the ability to recognize speech using a vocabulary of 1,011 words with reasonable accuracy. A significant contribution from the Harpy system was the introduction of a graph search concept, where language was represented as a connected network derived from lexical representations of words, incorporating syntactical production rules and word boundary rules.

In this system, the input speech underwent parametric analysis, followed by segmentation. The segmented parametric speech sequence was then subjected to phone template matching using the Itakura distance metric. The graph search, based on a beam search algorithm, compiled, hypothesized, pruned, and verified the recognized sequence of words or sounds that best satisfied knowledge constraints, achieving the highest matching score (smallest distance to the reference patterns). Notably, the Harpy system was among the first to leverage a finite state network to reduce computational load and efficiently identify the closest matching string.

However, this approach had its limitations, the primary one being the assumption that all phonemes had the same duration, which does not hold true. The subsequent project, also originating from CMU, aimed to address this specific problem.

HEARSAY-II

To address the aforementioned problem, Hearsay-II introduced what are known as symbolic problem solvers, each referred to as a 'knowledge source.' The need for multiple knowledge sources stemmed from the diverse transformations applied by speakers when creating acoustic signals and the corresponding inverse transformations required by listeners for interpretation[3].

These knowledge sources communicated through a blackboard architecture, which served as a repository for data, partial results, and finished conclusions. This design allowed each knowledge source to know where to retrieve information from the blackboard and where to post partial conclusions.

The range of knowledge sources included phonetic, phonemic, lexical, syntactic, semantic, prosodic, discursive, and even psychological aspects. Each of these sources could independently propose improved word string guesses for the given speech signal by 'writing' them on the virtual blackboard. Other knowledge sources could then build upon these suggestions.

Using this technique, Hearsay achieved recognition of 1011 words in continuous speech from multiple speakers with limited syntax, achieving an accuracy rate of approximately 90%. However, a notable limitation of this system was the time spent deciding which knowledge source to utilize next, which detracted from real-time speech processing—contrary to one of the project's primary goals. The subsequent system, Hear What I Mean (HWIM), aimed to address this challenge.

HWIM

Similar to Hearsay-II, the Hear What I Mean system was also knowledge-based, but it employed a more explicit scheduling approach based on human problem-solving methods. The process began with the identification of the most certain words, which served as 'islands of certainty,' and then leveraged these to iteratively expand both bottom-up and top-down identification processes.

Unlike using a central mechanism, in this system, knowledge sources would communicate with each other and share processed data. To score phoneme and word hypotheses, Bayesian probabilities were utilized. Furthermore, the system represented syntax using an Augmented Transition Network, presenting a greater challenge at the syntax level compared to previous systems.

Key Innovations

The outcome of this program was a system capable of accurately identifying 90%[4] of human-generated utterances from a vocabulary of 1000 words, a significantly larger vocabulary size than previous approaches.

Harpy achieved this by accessing word meanings from a database and determining sentence structure using its 'beam search' technology, a novel application of this approach. Furthermore, when Harpy encountered speech it couldn't understand, it responded with an 'I don't know what you said, please repeat' message, a feature reminiscent of today's voice assistants.

Impact

As a result of this project, it became evident that machines could be trained to comprehend not just individual words and numbers, but entire sentences. Consequently, subsequent research shifted its focus from discrete speech to continuous speech.

Future research

In this article, our primary focus has been on the successful or semi-successful outcomes of the project. However, it's important to note that some institutions, such as SDC, also participated in the project but did not achieve significant results. Therefore, it is worthwhile to shift our attention to their approaches and explore the reasons behind their lack of success.

References

  1. Gaon, A. (2021). The Future of Copyright in the Age of Artificial Intelligence. Великобритания: Edward Elgar Publishing.
  2. Funding a Revolution: Government Support for Computing Research. (1999). Украина: National Academies Press.
  3. The Hearsay II Speech-understanding System: Integrating Knowledge to Resolve Uncertainty. (1980). Соединенные Штаты Америки: ACM Press.
  4. Thorndyke, P. W., & Reddy, R. (1989, August). High-Impact Future Research Directions for Artificial Intelligence. In IJCAI (p. 1675).

Group members

  • Igor Marchenko
  • Wangyiyao Zhou
  • Yanpei Ouyang
  • Youyang Cai
  • Yi Lei