Patent Grant
7299173
1. Technical Field
The present invention relates to speech detection and, more particularly, relates to improved approaches to efficiently detect speech presence in a noisy environment by way of frequency and temporal considerations.
2. Description of the Related Art
In some applications, automatic speech recognition needs to be activated by uttering a particular word sequence such as keywords. For example, if a desktop personal computer has a speech recognizer for dictation or command control, it is desirable to activate the recognizer in the middle of the conversations in his or her office by uttering a keyword. This process of recognizing the keyword from continuous speech waveform is called keyword scanning. This would require the recognizer constantly recognizing the incoming speech and spotting those keywords. Nevertheless, the recognizer cannot be used to constantly monitor the incoming speech because it takes huge computational resources. Some other techniques that demand much less computations and memories have to be utilized to reduce the burden of speech recognizer. It is known that speech detection techniques are ways of eliminating silence segments from speech utterances so that speech recognizer can be speed up and do not wasting a lot of time on those silences or even misrecognize silence as speech. Speech detection techniques are often based on the speech waveform and utilize features such as short-time energy, zero crossing and etc. The same can be used to hypothesize keyword if some other features such as pitch, duration and voicing can be used in junction with word end-pointing techniques. Although the keyword hypothesis will be over generated, it still can reduce a large proportion of computations since the recognizer will only process these hypotheses.
Most speech recognition applications today face the challenging task of segmenting speech based on voice, unvoice & silence detection. A conventional approach is detecting short-term energy and zero crossings of a speech signal. These approaches are not reliable for noisy telephone speech signals due, in part, to the greater noise in a background environment of most telephone conversations. For example, stationary noise such as motor or wind noise and non-stationary noise such as door openings, closing or respiratory exhalation are present in telephone speech.
Accurate speech presence detection also conserves power and processing time for portable electronic devices such as cellular telephones. When reliable speech detection approaches are used, a speech recognition algorithm must find the utterances to determine if they are in fact language. This places a burden on computational complexity of processors and is a resource drain on portable electronic devices. A speech detection approach having computational efficiency as well as accuracy is needed.
The inventors of the present invention have discovered that there is a high variance associated with voiced speech such as vowels and the low variance associated with silences and wide-band noise. Speech presence can be efficiently detected in a noisy environment by way of frequency and temporal considerations using this variance.
Speech presence is detected by first bandpass filtering the speech to split it into banks of sub-bands. A matrix of shift registers secondly store each sub-band of speech. A power determining circuit then determines individual power measurements of the speech stored in each shift register element. A combining circuit combines the individual power measurements to provide a variance for the individual shift registers. A comparator circuit finally compares the variance with at least one threshold to indicate whether speech is detected. The present invention can be implemented by software in a microprocessor, digital signal processor or combinations with discrete components.
The details of the preferred embodiments of the invention will be readily understood from the following detailed description when read in conjunction with the accompanying drawings wherein:
Low band bandpass filter 141, mid band bandpass filter 143 and high band bandpass filter 145 split the preemphasized digital speech signal into a bank of preferably three sub-bands. Although a bank of three sub-bands is preferred, two or more sub-bands will work depending on the level of processing power and degree of detection accuracy needed for a noisy environment. It is preferred that the bandpass filters 141,143 and 145 divide the speech signal into somewhat equal sub-bands between 100 Hz and 3,000 Hz as follows. The low band bandpass filter 141 preferably has a band between 100 Hz and 1267 Hz, the mid and bandpass filter 143 preferably has a bandpass between 1267 Hz and 2433 Hz. The high band bandpass filter 145 preferably has a bandpass between 2433 Hz and 3600 Hz. Different band widths can be used for each sub-band.
A matrix of shift registers 150 receives the three sub-bands from the bandpass filters 141, 143 and 145. The shift registers 150 store each of the sub-bands and shifted to a next register location for each frame. In the preferred embodiment a total of three frames are stored in the shift registers, thus creating a three-by-three matrix Yij consisting of matrix elements Y11, Y12, Y13, Y21, Y22, Y23, Y31, Y32 and Y33. This matrix stores the speech information by way of both frequency and temporal considerations. Each of the three-by-three matrix elements contains sub-registers 250 for storing multiple samples k within a frame. For each of the register memories of the shift registers 150, a power measurement Xij is derived from the contents of the sub-registers. The calculation of the power measurements Xij for each sub-band over a frame i within a preferred 10 ms frame duration is performed by
The calculations of the power measurements Xij are preferably calculated within each of the matrix elements Yij of the shift register 150. The power measurement calculation sums the squares of each of the power samples for a particular sub-band over time. More detail for the preferred calculation of the power measurement for a sub-band across a number of samples in the shift register elements will later be described with reference to
The inventors of the present invention have discovered there is a high variance associated with voiced speech such as vowels and the low variance associated with silences and wide-band noise. A variance is a mathematical relationship known in digital speech processing as defined in elementary digital signal processing textbooks as such as Digital Communications, equations 1.1.65 or 1.1.66, by Proakis on page 17, published in 1989. The present invention applies a variance to a time-frequency power measurement to detect speech presence.
A variance combining circuit 160 calculates the variance of the plurality of power measurements for each sub-band and each frame. Calculating the variance VAR of the plurality of power measurements Xij for each sub-band j for each frame index i is calculated by
A comparator 170 compares the variance VAR with a threshold to determine whether or not the presence of speech is detected. When the variance is above the threshold, the presence of speech is detected, and a speech detection indication signal 180 is output. The threshold is preferably a fixed level however a variable threshold under certain conditions will yield more favorable results. A variable threshold can depend on determined by using an average of the past history of non-speech frames. Further, multiple thresholds can be implemented, one for clearly speech, one for clearly unspeech. A decision is made upon a transition over either of these thresholds.
The presence of speech indicated by the speech detection indication signal 180 can be used to gate on and off a speech recognition unit. The detection of the presence of speech is useful to gate and off a speech recognition unit so that the speech recognition unit does not need to operate continuously. This saves processing time that can be used for other purposes and/or conserves power, which reduces battery consumption in a portable electronic device. When a speech recognition circuit is present in a portable electronic device such as a cellular telephone, battery savings are achieved by freeing up the processor for other functions when speech presence is accurately determined. Also, the speech presence detection circuit does not require full activation of a recognition code so its more efficient. Reduction of miss-recognition is also achieved when using better speech presence accuracy. The speech detection indications are also useful for other devices such as speaker phones.
A power calculation circuit 259 calculates the average power among the sub-register elements for the given frame i and sub-band j. The average power Xij is calculated using the above equation (1). Each power calculation circuit 259 corresponds to one of the shift register elements in the matrix of
The signal processing techniques of the present invention disclosed herein with reference to the accompanying drawings are preferably implemented on one or more digital signal processors (DSPs) or other microprocessors. Nevertheless, such techniques could instead be implemented wholly or partially as discrete components. Further, it is appreciated by those of skill in the art that certain well known digital processing techniques are mathematically equivalent to one another and can be represented in different ways depending on the choice of implementation. For example the square of the terms in the variance calculation and/or power calculation can be substituted for absolute values without affecting the results.
Although the invention has been described and illustrated in the above description and drawings, it is understood that this description is by example only, and that numerous changes and modifications can be made by those skilled in the art without departing from the true spirit and scope of the invention. Although the examples in the drawings depict only example constructions and embodiments, alternate embodiments are available given the teachings of the present patent disclosure.
| Number | Name | Date | Kind |
|---|---|---|---|
| 4222115 | Cooper et al. | Sep 1980 | A |
| 4461024 | Rengger et al. | Jul 1984 | A |
| 4827519 | Fujimoto et al. | May 1989 | A |
| 5097510 | Graupe | Mar 1992 | A |
| 5617508 | Reaves | Apr 1997 | A |
| 5659622 | Ashley | Aug 1997 | A |
| 5692104 | Chow et al. | Nov 1997 | A |
| 5732392 | Mizuno et al. | Mar 1998 | A |
| 5826230 | Reaves | Oct 1998 | A |
| 5963901 | Vahatalo et al. | Oct 1999 | A |
| 5991718 | Malah | Nov 1999 | A |
| 6278972 | Bi et al. | Aug 2001 | B1 |
| 6397050 | Peterson et al. | May 2002 | B1 |
| 6591234 | Chandran et al. | Jul 2003 | B1 |
| 6711536 | Rees | Mar 2004 | B2 |
| Number | Date | Country |
|---|---|---|
| 0 945 854 | Sep 1999 | EP |
| JP9401181 | Feb 1996 | WO |
| WO 0111606 | Feb 2001 | WO |
| Number | Date | Country | |
|---|---|---|---|
| 20030144840 A1 | Jul 2003 | US |
