Frequently asked questionsGeneralTechnicalBusiness
General
To answer this question, we have to answer some questions of our own. Why is it that we human beings spend 10-12 years studying grammar in school, yet until today, its been mostly ignored in speech recognition by computers? Although some research has been done to introduce syntactic related rules into speech recognition with limited success, to the best of our knowledge, prior research has not used syntax as a building block to get to recognize concepts. Communicating concepts is the main thing that matters in human speech; the rest is just instrumental. So then, why shouldn’t a computer process speech in the same way? Let us explain. The human mind does not recognize phonemes. It is listening for “concepts”. It hears a sound and looks for a means to understand what it represents. In most cases the sounds it hears are not complete enough to identify the words in a sequence needed to form a concept. Some researchers say “we fill-in the missing data” but this is not a totally accurate statement. We, human beings, evaluate the sounds we receive in several ways. First, we listen for the words that could be represented by the sounds. Second, we evaluate these “words” based on their sequence (syntax) and the subject (conceptual-domain) being addressed. Nevertheless, getting a correct syntactic organization is not sufficient in speech. Think about the sequence "Red rounded house eats a squared ship". This is a syntactically well built sequence, but it does not make any sense. If the words sequence have a valid syntax and match the subject, then we try to derive a concept from it. The process is repeated until a valid concept is detected. For a human brain, this is instantaneous and subconscious. We must take into consideration, that the computing power of the human mind is virtually unlimited. Now, how do we get to the same result using only the limited power of today’s computers? Research has determined that a spoken language can be broken-down into sounds representation called phonemes. When today’s speech recognition systems analyze sound they identify the phonemes and use them to builds words. Accuracy of theses systems, as the ones used by state of the art speech recognition, is dependent on identifying the correct phoneme every time (forcing them to limit their vocabulary). Conceptual Speech Recognition does not process sound in such a way. In our approach, when a computer hears "this is a test", it may not identify immediately the right phoneme for each time slice. The 't' phoneme of 'test' also probably had a reasonable probability of being recognized as a 'p' phoneme based on sound alone. Then, 'pest' may as well have been recognized as 'test'. We keep track of several potential phonemes for each time slice. Then we generate multiple potentially spoken words derived from these potential phonemes recognized. The statement “this is a test” will produce many potential words. Using all these potential words, we evaluate them in combinations, looking for syntactically correct sequences. Once a syntactically valid sequence is identified, we assess its conceptual content based on the domain (subject) being addressed. We continue to recombine the potential words together looking for valid syntax and conceptual content until it succeeds. Processing speech in such a manner results in exceptional accuracy. Using this approach we have introduced the syntactic and conceptual aspect of speech into computer based speech recognition processing. Although there have been previous attempts to introduce syntactic related rules into speech recognition, the success of these approaches has been limited because syntax is only a building block to get to concepts. Concepts are the only thing that matter in speech; the rest is just instrumental.
Our technology frees speakers from the endless array of voice prompts and regimented commands that burden many of today's IVR systems. With conceptual speech recognition, the speaker is freed from having to speak commands in a predetermined way. As long as the speaker conveys the general idea in his or her statement, conceptual speech recognition can determine the concept and execute the command. We do this by looking for the concept being spoken as it relates to the subject matter. In the telephony environment where the sound sampling is about 8KHz, our technology works without the use of navigation menus. End users are free to state their request in their own words, and receive a meaningful reply or be placed directly at the point in the system where they need to be. This takes place in one simple step without having to follow a series prompts. Future applications of this technology will prove equally efficient for desktop applications and other implementations where higher sound sampling is available. This is possible because our technology relies more on conceptual analysis, than on sound recognition. Humans can communicate complex thoughts over a simple telephone line because they use concepts to fill in the gaps when sounds are misperceived. The sounds we hear build potential words that we link together to build concepts, and we pick the single most probable concept as the recognized one. Similar to the human mind, our system looks for concepts related to the subject matter, and selects the best concept within the scope of the subject matter.
The primary customers for this technology are IVR development and deployment companies that provide command and control telephony solutions to the market. Most of today's IVR systems that rely heavily on voice prompts or scripted commands are good candidates for integrating conceptual speech recognition. The result will be a more positive experience for the end user and a reduced opt-out rate. Eventually, conceptual speech recognition will open new markets in the speech recognition industry. Imagine a drive-thru ordering system for a fast food restaurant, as an example, where a client will be able to place an order by saying "I want a double cheese-burger without onions and (Eh) put extra mustard on it and make this a combo". These types of utterances that fall within a specific domain are easily handled by conceptual speech recognition and will produce the expected response. Many things we do with computers are command and control, and our technology will expand the usefulness of speech recognition to a wider range of tasks. The result will be increased productivity and a more positive experience for end users.
"Natural Language understanding" (NLU) or "Natural Language Modeling" (NLM) is a terminology commonly used in speech recognition to describe a means for performing speech recognition in which the user can say commands naturally, as if speaking to another human being. "Conceptual Dependency" (CD) is an application of "Knowledge Representation" research in the field of natural language understanding and philosophy. It is a relatively new science that was born in the 1970's. It involves having a computer perform the parsing of text in a specific manner to conceptually represent meaning structures and ideas without regard to how the words were initially organized syntactically.
Most current speech recognition products use some form of natural language modeling. However, their usefulness is limited by the scope of the NLM and the quality of the speech recognition. Most of these systems do not truly allow users to speak naturally or issue commands in an unrestricted way. In short, you can speak naturally, as long as the system recognizes your speech and that you limit your speech to the confines of what is defined in the natural language model. Conceptual speech recognition looks for concepts in the spoken command, and as long as the user conveys the general concept (or configurations of multiple concepts that are handled), the system can generate the appropriate response. The system is no longer restricted by NLM, and the user is truly free to speak naturally, just as they would to another human being.
A conceptual dependency parser is a process that extracts syntactic related information from language in order to expose the concept in a pure and standardized expression. Syntactic analysis is one of the two important steps in order to produce a valid conceptual representation. A conceptual dependency parser first needs to extract the syntactic organization of the language to be parsed, and then apply rules related to concept building associated with each word (while taking into consideration the syntactic hierarchy and the nature of each word - verb, noun, preposition, etc) in order to produce a final conceptual representation.
Using conceptual speech recognition frees speakers from an endless series of prompts and limitations of predefined commands. End users are able to speak freely in their own words when making an inquiry or issuing a command. State of the art speech recognition engines used in IVR systems are entirely dependent on sound to drive their process. But, because of the imperfect nature of sound that is affected by noise and pronunciation variations, IVR system developers are required to restrict requests to a limited set by using menu structures that artificially improves accuracy at every step. Because our system relies on conceptual analysis instead of solely sound matching, accuracy is enhanced (even over an average telephone line bandwidth of 8 kHz). Using conceptual speech recognition frees developers from the limitations of predefined commands. This enhances the user experience in most IVR applications and reduces the opt-out rate. Finally, the technology enables a speaker to ask for multiple things in a single inquiry and still get a correct response for each part.
All those objectives are attained through a process that is extremely efficient (often analyzing complex utterances in less than 1/4 of a second), and that does not require more effort to produce than current state of the art speech recognition. The result will be a more positive end-user experience. Call times will be reduced significantly since speakers will be able to get right to the point and get a correct response without having to navigate complex and often frustrating menus where they would normally bypass the system to talk to a live person.
Absolutely! Conceptual speech recognition has been in research and development for three years and strong demonstrations as well as working examples using the approach are already available. An SDK is currently being developed. This SDK will enable system developers to produce conceptual speech recognition systems or integrate the technology into existing systems with minimal time and expense. In addition, we filed our patent application in June 2003 and our technology is available for licensing. Detailed technical information on the technology is available on our website. Using this information, engineers can review our technology and make their own determinations. There is enough information for engineers to review our technology and make their own determinations. We are confident that our technology will withstand scrutiny and is indeed a better way to handle speech recognition. In summary, the market is ready for conceptual speech recognition, and the conceptual speech recognition is ready for the market.
The time and effort required to implement conceptual speech recognition in real-life applications is comparable to current state of the art speech recognition projects. Although initial conceptual speech recognition projects may require more time to complete, support will be available and it will not take long for a qualified development team to ramp up their production. Conceptual speech recognition introduces some new concepts in order to produce solutions. Those new concepts are limited and relatively simple to understand and implement by an engineer.
People implementing conceptual speech recognition systems need to be experienced in C/C++ coding. They also need a superficial understanding of some linguistics concepts. A superficial understanding of parts of speech that are associated with each word is sufficient (verbs, nouns, prepositions, noun phrases, etc). They will also be required to produce what is called 'Predicate Builder Scripts' in order to build conceptual representations from spellings for the conceptual domain of the application. This new technology was introduced by Conceptual Speech LLC in its June 30, 2003 patent application where it is well documented. Our SDK, Conceptual Speech Commander, will also include the necessary documentation and support needed to build/integrate the technology into an application. An experienced engineer will not have any difficulty understanding and implementing this logic.
Yes, there is a working demonstration of the technology that simulates a real-world application in an airline response system that produces responses to telephone inquires related to flight scheduling. Documentation of this demonstration is available on our website by going here. We are also glad to show potential customers a live demonstration of our demo. Please contact us at sales@conceptualspeech.com for more information.
Refer to Conceptual Speech LLC's web site here to get access to further documentation on conceptual speech recognition. These pages contain detailed information on the technology.
Using conceptual speech recognition, you will be able to enable end users of your IVR systems to quickly and efficiently navigate to the point where they need to be in the system. This will significantly reduce the number of callers that opt-out of the system out of frustration with an endless chain of voice prompts that waste time and money. For example, the following is a hypothetical scenario of a customer calling the Phone Company. System: "Welcome. What can I do for you?" Caller: "I’d like to pay my phone bill with a credit card." System: "Certainly. Your call is now being transferred to our payment system. Thank you." The caller is quickly transferred to the system prompts where the payment amount and credit card information can be provided. However, the caller could as well have responded to the initial prompt in a different way. System: "Welcome. What can I do for you?" Caller: "I’d like to add caller ID to my phone service." System: "Certainly. Your call is now being transferred to a sales representative. Thank you." Technical
Conceptual Speech LLC is hard at work moving towards completion of the first expression of this revolutionary technology. The name of the product is set to be Conceptual Speech Commander. Conceptual Speech Commander is a development and deployment toolkit designed for use in implementing command and control conceptual speech recognition. Initial deployment of conceptual speech recognition will be designed for command and control environments where the vocabulary is well defined - for example - an airline response system or a fast food ordering system where the conceptual domain is well defined and the speaker is not expected to ask about the weather or sports scores. Please contact us to learn more about the Conceptual Speech Commander and find out how implementing conceptual speech recognition can enhance customer experience in your applications. Email sales@conceptualspeech.com.
Are you able to extract without any bias from your own sound acquisition a finite set of probable phonemes for each time-slice where the right phoneme will always be in that set? If you are able to do that, and need an average of 2 phonemes from a unique cluster per time-slice, performance will be excellent. If an average of more than 4 phonemes from a unique cluster per time-slice were required, performance would start to decrease. If such result is not reachable because some phonemes were not detected successfully, it is still possible to use Conceptual Speech Commander. A skipped phoneme error correction is indeed available in Conceptual Speech Commander'. The most important rule related to error correction is that two consecutive errors are not permitted. That is, if one phoneme is skipped in your phoneme streaming, there is no problem as long as the next phoneme that should have been recognized is not also skipped. Enabling error correction will affect performance and a phoneme streaming process that does not require error correction is preferable.
Conceptual dependency is well known to have some limitations. More specifically, three major limitations are often associated with conceptual dependency.
We, at Conceptual Speech LLC, are well aware of CDs limitations. On the other hand, we are also well aware of its strengths. We concentrated on CDs strengths while securing the fact that its supposed weaknesses are not an obstacle to creating a useful speech recognition approach. We did this the following way.
Avoiding conceptual dependency's limitations while using its numerous strengths allowed the development of conceptual speech recognition. We see this technology as becoming an important tool in speech recognition in the immediate future. Our approach to resolving the complex problem of recognizing speech is solid, our implementation in Conceptual Speech Commander is well thought and its deployment will be a major event in the evolution of speech recognition.
Our SDK already comes with a huge vocabulary of 250 000 words and associated pronunciations and parts of speech required to build a conceptual speech recognition system. It also includes a syntactic and conceptual analyzers that are ready for the English language. Syntactic rules required to process English content are available with the SDK and, if you feel some rules were left out of it, you can add to it. The SDK also comes with a complete testing environment that will help you test your system prior and during deployment by producing detailed logs on how utterances are processed during all phases of processing (phoneme analysis, phoneme stream analysis, syntactic analysis and conceptual analysis). When deployed, the system may be unable to handle some calls because of vocabulary used that were unexpected. For those calls, the system can simply transfer to an operator. A detailed log file of each failed call is generated on the system and kept for later referral from one of your engineers so that the system can be improved. He will even be able to reprocess an audio file obtained from the original failed call until he gets the system to handle it properly. Over time, your system will improve so much that such occurrences will rarely happen. A powerful language is also available from the SDK to define words conceptually. That language is called Predicate Builder. His sole purpose is to provide the conceptual rules associated with each word required in order to build concepts (predicates). The language is simple to understand and easy to program for an engineer skilled in the art. Business
We, at Conceptual Speech Technologies LLC, are committed to two goals:
| |||||||||||||||||||||||||||||||||||||
|
