Distributed speech recognition method

#The subject application is a U.S. National Stage application that claims the priority of International Application No. PCT/FR2004/00547, filed on 8 Mar. 2004, which claims the priority of French National Application No. 03 03614, filed on 25 Mar. 2003.

#The present invention relates to the field of the vocal control of applications, practiced on user terminals, thanks to the implementation of speech recognition means. The user terminals in question are all devices with a means of sensing speech, commonly a microphone, having capabilities for processing this sound and being connected to one or more servers via a transmission channel. These can, for example, be control or remote control devices used in home office applications, in automobiles (car radio or other vehicle function control), in PCs or telephones. The range of applications concerned is essentially that where the user controls an action, requests information or wants to interact remotely using a voice command. The use of voice commands does not exclude the existence within the user terminal of other activation means (multi-modal system), and the return of information, system states or responses may also be effected in visual, audio, olfactory or any other humanly perceptible combined form.

#Generally speaking, the means for effecting speech recognition comprise means for obtaining an audio signal, means for acoustic analysis which extract modeling parameters and, finally, recognition means that compare these extracted modeling parameters with models and suggest the form stored in the models that can be associated with the signal in the most probable manner. Optionally, voice activation detection (VAD) means may be used. These provide the detection of the sequences corresponding to speech which are required to be recognized. They extract segments of speech from the audio signal at the input, outside of the periods without voice activity, which will subsequently be processed by the modeling parameter extraction means.

#The system shown in the single FIGURE comprises a server 1 and a user terminal 2, which communicate with one another via a network (not shown) having channels for the transmission of voice signals and channels for the transmission of data signals.

#The terminal 2 comprises a microphone 4 which collects the speech to be recognized from a user in the form of an audio signal. The terminal 2 also comprises modules 5, 6, 7. The voice activation detector module VAD 5 provides the detection of the sequences corresponding to speech and which are required to be recognized. This module 5 is for example designed for fast detection of command words. The module 6 performs an acoustic analysis in a manner known per se: it calculates modeling parameters, while at the same time providing a noise-filtering function.

#The module 7 executes a recognition algorithm of a known type, for example based on hidden Markov models with a reduced vocabulary. This recognition engine 7 can operate in mono-speaker mode and requires an initial learn phase for the voice of the user.