Improving speech recognition of mobile devices

Description of equivalent WO 2004001719 A1

Improving Speech Recognition Of Mobile Devices

Background

This invention relates generally to mobile processor-based systems that include speech recognition capabilities.

Mobile processor-based systems include devices such as handheld devices, personal digital assistants, digital cameras, laptop computers, data input devices, data collection devices, remote control units, voice recorders, and cellular telephones, to mention a few examples. Many of these devices may include speech recognition capabilities.

With speech recognition, the user may say words that may be converted to text. As another example, the spoken words may be received as commands that enable selection and operation of the processor-based system's capabilities.

In a number of cases, the ability of a given device to recognize speech or identify a speaker is relatively limited. A variety of ambient conditions may adversely affect the quality of the speech recognition or speaker identification. Because the ambient conditions may change unpredictably, the elimination of ambient effects is much more difficult with mobile speech recognition platforms.

Thus, there is a need for better ways to enable speech recognition with mobile processor-based systems.

Brief Description of the Drawings Figure 1 is a schematic depiction of one embodiment of the present invention;

Figure 2 is a flow chart useful with the embodiment shown in Figure 1 in accordance with one embodiment of the present invention; and

Figure 3 is a flow chart useful with the embodiment shown in Figure 1 in accordance with one embodiment of the present invention.

Detailed Description

Referring to Figure 1, a speech enabled mobile processor-based system 14 may be any one of a variety of mobile processor-based systems that generally are battery powered. Examples of such devices include laptop computers, personal digital assistants, cellular telephones, digital cameras, data input devices, data collection devices, appliances, and voice recorders, to mention a few examples.

By incorporating a position detection capability within the device 14, the ability to recognize spoken words may be improved in the variety of environments or ambient conditions. Thus, the device 14 may include a position detection or location-based services (LBS) client 26. Position detection may be accomplished using a variety of technologies such as global positioning satellites, hot-spot detection, cell detection, radio triangulation, or other techniques.

A variety of aspects of location may be used to improve speech recognition. The physical location of the system 14 may provide information about acoustic characteristics of the surrounding space. Those characteristics may include the size of the room, noise sources, such as ventilation ducts or exterior windows, and reverberation characteristics.

This data can be stored in a network infrastructure, such as a location-based services (LBS) server 12. For frequently visited locations, the characteristics may be stored in the system 14 data store 28 itself. The server 12 may be coupled to the system

14 through a wireless network 18 in one embodiment of the present invention.

Other aspects of location that may be leveraged to improve speech recognition include the physical location of nearby speakers who are using comparable systems 14.

These speakers may be potential sources of interference and can be identified based on their proximity to the user of the system 14. In addition, the identity of nearby people who are carrying comparable systems 14 may be inferred by subscribing to their presence information or by ad hoc discovery peers. Also, the orientation of the system 14 may be determined and this may provide useful information for improving speech recognition.

The system 14 includes a speech context manager 24 that is coupled to the position detection/location-based services client 26, a speech recognizer 22, and a noise mitigating speech preprocessor 20.

When speech recognition is attempted by the system 14, the speech context manager 24 retrieves a current context from the server 12 in accordance with one embodiment of the present invention. Based on the size of the surrounding space, the context manager 24 adjusts the acoustic models of the recognizer 22 to account for reverberation. This adjustment may be done in a variety of ways including using model adaptation, such as maximum likelihood linear regression to a known target. The target transformation may have been estimated in a previous encounter at that position or may be inferred from the reverberation time associated with the space. The adjustment may also be done by selecting from a set of previously trained acoustic models that match various acoustic spaces typically encountered by the user.

As another alternative, the context manager 24 may select from among feature extraction and noise reduction algorithms that are resistant to reverberation based on the size of the acoustic space. The acoustic models may also be modified to match the selected front-end noise reduction and feature extraction. Models may also be adapted based on the identity of nearby people, retrieving and loading speaker dependent acoustic models for each person, if available. Those models may be used for automatic transcription of hallway discussion in one embodiment of the present invention.

Another way that the adjustment may be done is by initializing and adapting a new acoustic model if the acoustic space has not been encountered previously. Once the location is adequately modeled, the system 14 may send the information to the server 12 to be stored in the remote data store 16 for future visitors to the same location.

As another example of adaptation, based on the identity of nearby speakers, the system 14 may assist the user to identify them as a transcription source. A transcription source is someone whose speech should be transcribed. A list of potential sources in the vicinity of the user may be presented to the user. The user may select the desired transcription sources from the list in one embodiment.

As still another example, based on the orientation of the system 10, the location of proximate people, and their designation as transcription sources, a microphone array controlled by preprocessor 20 may be configured to place nulls in the direction of the closest persons who are not transcription sources. Since that direction may not be highly accurate and is subject to abrupt change, this method may not supplant interferer tracking via a microphone array. However, it may provide a mechanism to place the nulls when the interferer is not speaking, thereby significantly improving performance when an interferer talker starts to speak.

Referring to Figure 2, in accordance with one embodiment of the present invention, the speech context manager 24 may be a processor-based device including both a processor and storage for storing instructions to be executed on the processor. Thus, the speech context manager 24 may be software or hardware. Initially, the speech context manager 24 retrieves a current context from the server 12, as indicated in block 30. Then the context manager 24 may determine the size of the surrounding space proximate to the device 14, as indicated in block 32. The device 14 may adjust the recognizer's acoustic models to account for local reverberation, as indicated in block 34.

Then feature extraction and noise reduction algorithms may be selected based on the understanding of the local environment, as indicated in block 36. In addition, the speaker-dependent acoustic models for nearby speakers may be retrieved and loaded, as indicated in block 38. These models may be retrieved, in one embodiment, from the server 12.

New acoustic models may be developed based on the position of the system 14 as detected by the position detection/LBS client 26, as indicated in block 40. The new model, linked to position coordinates, may be sent over the wireless network 18 to the server 12, as indicated in block 42, for potential future use. h some embodiments, models may be available from the server 12 and, in other situations, those models may be developed by a system 14 either on its own or in cooperation with the server 12 for immediate dynamic use.

As indicated in block 44, any speakers whose speech should be recognized may be identified. The microphone array preprocessor 20 may be configured, as indicated in block 46. Then speech recognition may be implemented, as indicated in block 48, having obtained the benefit of the location information.

Referring to Figure 3, the LBS server 12 may be implemented through software 50 in accordance with one embodiment of the present invention. The software 50 may be stored in an appropriate storage on the server 12. hiitially, the server 12 receives a request for context information from a system 14, as determined in diamond 52. Once received, the server 12 obtains the location information from the system 14, as indicated in block 54. The location information may then be correlated to available models in the data storage 16, as indicated in block 56. Once an appropriate model is identified, the context maybe transmitted to the device 14 over the wireless network, as indicated in block 58.

While the present invention has been described with respect to a limited number of embodiments, those skilled in the art will appreciate numerous modifications and variations therefrom. It is intended that the appended claims cover all such modifications and variations as fall within the true spirit and scope of this present invention.