1. Field of the Invention
The present invention relates to spoken dialog systems and more specifically to system and method of building application-dependent text-to-speech custom voices.
2. Introduction
State-of-the-art spoken dialog systems include several components that enable the system to understand speech spoken by a user, generate a meaningful response, and then audibly speak the response. These basic components of such a system 100 are shown in
The method for generating a TTS voice usually involves a costly and time-consuming process. For example, if a person desires to have their voice be used for a TTS voice in a spoken dialog system, several steps are typically necessary to build the custom voice. First, a developer selects text material for reading by the person. The text may relate to a specific domain of the spoken dialog system. An example of such a system may be a travel reservation system. The person would then be given text that relates to the context or domain of travel reservations, i.e., “what is your destination city?” The process of creating the custom voice then involves recording a speech corpus of the person to obtain data from which to generate the custom voice. This typically involves recording 10-20 hours of the person speaking or reading the selected text, and processing the speech to obtain an inventory of speech units that can be concatenated together to create a TTS voice. This is a very computationally intensive process and a time consuming process. For example, the time to build such a custom voice may take a month or more. In addition, the human expertise and professional interaction necessary to build such a custom voice is high. Significant human effort is required to create the custom voice.
The cost for such a process is prohibitive as well. The high cost to a potential buyer of a custom voice to collect the speech, label the speech and build the custom voice using the above-described approach prevents many companies from deploying a spoken dialog service. To accommodate for this cost, some companies use recorded prompts in a spoken dialog system. This approach, however, dramatically limits the flexibility and adaptability of the spoken dialog service to new questions and new interactions with the users. The cost to record enough prompts to handle every scenario also becomes time consuming and cost prohibitive.
What is needed in the art is a more efficient and less expensive approach to generating a custom, in-domain TTS voice.
Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The features and advantages of the invention may be realized and obtained by means of the instruments and combinations particularly pointed out in the appended claims. These and other features of the present invention will become more fully apparent from the following description and appended claims, or may be learned by the practice of the invention as set forth herein.
The present invention provides a method for creating high-quality customized application voices automatically with zero or little human interaction. The method requires no application speech data to be collected. One aspect of the invention relates to a method of generating a custom text-to-speech voice. The method comprises selecting a voice for generating a custom text-to-speech voice associated with an application domain, and collecting text data associated with the domain from a pre-existing text data source. Using the collected text data, the method further comprises generating an in-domain inventory of synthesis speech units by selecting speech units appropriate to the domain via a search of a pre-existing much bigger inventory of synthesis speech units and generating a text-to-speech custom voice for the domain utilizing the generated in-domain inventory of synthesis speech units.
To further enhance the quality of the TTS voice, active learning techniques may be employed to identify perceptually problematic sentences, phrases, words or speech units. From these identified problems, the developer can record speech associated with the problem areas and label the data to integrate that data into the in-domain inventory of speech units. In this manner, the quality of the synthesized voice is enhanced with only a few additional minutes of recorded data. Other embodiments of the invention are more fully described below.
In order to describe the manner in which the above-recited and other advantages and features of the invention can be obtained, a more particular description of the invention briefly described above will be rendered by reference to specific embodiments thereof which are illustrated in the appended drawings. Understanding that these drawings depict only typical embodiments of the invention and are not therefore to be considered to be limiting of its scope, the invention will be described and explained with additional specificity and detail through the use of the accompanying drawings in which:
Various embodiments of the invention are discussed in detail below. While specific implementations are discussed, it should be understood that this is done for illustration purposes only. A person skilled in the relevant art will recognize that other components and configurations may be used without parting from the spirit and scope of the invention.
The present invention utilizes data 210 to automate the process of creating a synthetic TTS custom voice. While the preferred embodiment fully automates this process, there are alternate approaches where a minor portion of the process requires human intervention as will be seen from the description below.
A preliminary step in an aspect of the present invention is inventory preparation. To automate the voice generation process, first an inventory of speech units needs to be prepared.
There are several ways a custom voice can be generated in this situation. One exemplary approach is to represent the models and data as finite state automata and apply statistical adaptation algorithms to bias the domain specific data. The resulting model not only promises to provide better quality output speech but also speed-ups the processing time. Given that these operations are essentially finite state compositions, they can be conducted within minutes. As task-specific speech data becomes available (through active learning or other methods), the same paradigm can be followed either in a supervised or in an unsupervised manner.
To further improve quality, active learning can be applied for unit selection so as to automatically identify poor quality data and to identify specific areas where speech recordings should occur in preparing the TTS voice. Active learning can enable a new system to be integrated into an existing system easily and inexpensively. Thus, the present invention can enhance the synthesis quality with only the minimum set of recorded data using the principles of active learning. The details and the various algorithms utilized in active learning are known to those of skill in the art and thus further information is not provided herein otherwise to note that this technology may be utilized as a component to automatically generating a TTS voice.
The approach of the present invention provides many advantages. For example, the quality of the TTS voice can be close to that of natural speech and can match the quality of recorded prompts if special-domain voices are used. There are cost savings associated with an automated or nearly automated process over the conventional systems. Another advantage is that updated information, changes in data, or changes in the subject matter of the domain can automatically be integrated into the system and a new TTS voice automatically deployed covering the new or changed subject matter. This is essential for up-to-date information rendering where information related to the domain is dynamic. This is also useful for personalization of the TTS voice.
Another module can then automatically extract relevant detailed text data and phrases that are salient to the application being developed. A variety of approaches may be used to identify the salient phrases including n-gram selection, maximal mutual information and other means. These phrases (speech units, words, phrases, sentences or paragraphs) are extracted, pre-processed and added to the inventory as part of the search strategy. Another approach is to add salient phrases manually to the inventory database. Using the collected text data, an in-domain inventory of synthesis speech units is generated by selecting speech units appropriate to the domain via a search of a pre-existing inventory of synthesis speech units (506). The text-to-speech custom voice is generated for the domain utilizing the in-domain inventory of synthesis speech units (508). As mentioned above, active learning may be applied to identify problem phrases (510) and based on the identified problem phrases, one can record and prepare data for the in-domain inventory according to the identified problem phrases (512). The method further comprises determining whether the quality of the synthesis is sufficiently high (514). If not, then the process returns to step 504 for further collection of data and to refine and improve the quality. In this regard, an alternate approach is to determine a minimal inventory for a selected level of synthesis quality for collection of data in step 504 and/or recording a minimal inventory in step 512. This reduces the workload at various stages in the process according to a selected level of quality. Then, if the quality of the synthesized voice is not sufficiently high in step 514, then more data collection occurs and/or more recording occurs in an iterative fashion until the requisite quality level is met. It may also be determined after step 514 that the only further preparation needed is the recording of more inventory, in which case the process would proceed from step 514 directly to step 512 for more recording followed by another quality check. This leads to a method where the process of creating a voice can be iterated until a sufficiently high quality synthesis has been achieved and where the synthesis generation process more closely matches the chosen synthesis quality.
Another example of the principles of the invention in operation would be for web interaction. Suppose a website has a “talking head” or 3D animated head that guides a person through the website. The service provider or talking head provider may enable the user to select various characters or features for the animated head. One of the options may be a voice option. If the user selects to use his or her own voice for interaction with the website, then the training process according to the present invention could utilize the existing text on that website for building the new customer voice for the user. Such a website could also have a text database of existing salient phrases such that database searches outside the scope of the website are unnecessary. The user may provide his own voice via a microphone connected to his computer or via a telephone connection. Text can be presented via the website for the user to read to collect voice segments for the inventory of speech units. The user may even approve a phone company recording over a period of time of his or her phone conversations to obtain hours of recordings of his or her voice. Other parties to the conversation would have any recording of their voices deleted. In this regard, the TTS voice may not necessarily be incorporated into a spoken dialog system where the user becomes engaged in a conversation with the animation. Any kind of appliance where a voice or a talking head animation would provide added value could utilize the principles of this invention. Thus, other embodiments of the invention relate to the various websites, handheld computing devices, kiosks, or any appliance employing a TTS voice customized according to the invention disclosed herein.
Embodiments within the scope of the present invention may also include computer-readable media for carrying or having computer-executable instructions or data structures stored thereon. Such computer-readable media can be any available media that can be accessed by a general purpose or special purpose computer. By way of example, and not limitation, such computer-readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to carry or store desired program code means in the form of computer-executable instructions or data structures. When information is transferred or provided over a network or another communications connection (either hardwired, wireless, or combination thereof) to a computer, the computer properly views the connection as a computer-readable medium. Thus, any such connection is properly termed a computer-readable medium. Combinations of the above should also be included within the scope of the computer-readable media.
Computer-executable instructions include, for example, instructions and data which cause a general purpose computer, special purpose computer, or special purpose processing device to perform a certain function or group of functions. Computer-executable instructions also include program modules that are executed by computers in stand-alone or network environments. Generally, program modules include routines, programs, objects, components, and data structures, etc. that perform particular tasks or implement particular abstract data types. Computer-executable instructions, associated data structures, and program modules represent examples of the program code means for executing steps of the methods disclosed herein. The particular sequence of such executable instructions or associated data structures represents examples of corresponding acts for implementing the functions described in such steps.
Those of skill in the art will appreciate that other embodiments of the invention may be practiced in network computing environments with many types of computer system configurations, including personal computers, hand-held devices, multi-processor systems, microprocessor-based or programmable consumer electronics, network PCs, minicomputers, mainframe computers, and the like. Embodiments may also be practiced in distributed computing environments where tasks are performed by local and remote processing devices that are linked (either by hardwired links, wireless links, or by a combination thereof) through a communications network. In a distributed computing environment, program modules may be located in both local and remote memory storage devices.
Although the above description may contain specific details, they should not be construed as limiting the claims in any way. Other configurations of the described embodiments of the invention are part of the scope of this invention. For example, a car or vehicle may communicate with the driver via a personalized custom voice. In this regard, the invention may be a vehicle having a custom personalized voice. There are many products and services that may incorporate a customized voice generated according to the present invention. Accordingly, the appended claims and their legal equivalents should only define the invention, rather than any specific examples given.
Number | Name | Date | Kind |
---|---|---|---|
5396577 | Oikawa et al. | Mar 1995 | A |
6601030 | Syrdal | Jul 2003 | B2 |
7219060 | Coorman et al. | May 2007 | B2 |
7328157 | Chu et al. | Feb 2008 | B1 |
20030216921 | Bao et al. | Nov 2003 | A1 |
20040006471 | Chiu | Jan 2004 | A1 |
20040111271 | Tischer | Jun 2004 | A1 |
20040254792 | Busayapongchai et al. | Dec 2004 | A1 |
20080114756 | Konig et al. | May 2008 | A1 |
Number | Date | Country | |
---|---|---|---|
20050256716 A1 | Nov 2005 | US |