Automatic utterance detector with high noise immunity

[0001] This application claims priority under 35 USC § 119(e)(1) of provisional application No. 60/161,179, filed Oct. 22, 1999.

[0002] This invention relates to speech recognition and, more particularly, to an utterance detector with high noise immunity for speech recognition.

[0003] Typical speech recognizers require an utterance detector to indicate where to start and to stop the recognition of the incoming speech stream. Most utterance detectors use signal energy as basic speech indicator. See, for example, J.-C. Junqua, B. Mak, and B. Reaves, “A robust algorithm for word boundary detection in the presence of noise,” IEEE Trans. on Speech and Audio Processing, 2(3):406–412, July 1994 and L. Lamels, L. Rabiner, A. Rosenberg, and J. Wilpon, “An improved endpoint detector for isolated word recognition,” IEEE ASSP Mag., 29:777–785, 1981.

[0004] In applications such as hands-free speech recognition in a car driven on a highway, the signal-to-noise ratio can be less than 0 db. That means that the energy of noise is about the same as that of the signal. Obviously, while speech energy gives good results for clean to moderately noisy speech, it is not adequate for reliable detection under such a noisy situation.

[0005] In accordance with one embodiment of the present invention, an utterance detector with enhanced noise robustness is provided. The detector is composed of two components: frame-level speech/non-speech decision and utterance-level detector responsive to a series of speech/non-speech decisions.

[0022] Referring to FIG. 1, there is illustrated a block diagram of the utterance detector 10 according to one embodiment of the present invention. The detector 10 comprises the first part which is at the frame level detector 11 which determines for each frame if there is speech or non-speech. The second part is an utterance detector 13 that includes a state machine that determines if the utterance is speech. The output of the utterance detector 13 is applied to speech recognizer 16 such that when the utterance detector recognizes speech it enables the recognizer 16 to receive speech and when the detector determines non-speech to turn off or disable the recognizer 16.

[0023] FIG. 2 illustrates the system. Row (a) of FIG. 2 illustrates a series of frames 15. In the first detector 11, it is determined if the frame 15 is speech or non-speech. This is represented by row (b) of FIG. 2. Row (c) of FIG. 2 represents the utterance decision. Detected speech in a frame at frame detector 11 causes the higher level signal and the low or lower level signals for each frame is represented by the lower level signal. In the utterance decision, only after a series of detected speech frames does the utterance detector 13 enable the recognizer.

[0024] In the prior art, energy level is used to determine if the input frame is speech. This is not reliable since noise such as highway noise could have as much energy as speech.

[0025] For resistance to noise, Applicants teach to exploit the periodicity, rather than energy, of the speech signal. Specifically, we use autocorrelation function. The autocorrelation function (correlation with signal delayed by τ) used in this work is derived from speech X(t), and is defined as:
R_x(τ)=E[X(t)X(t+τ)]  (1)

[0026] Important properties of R_x(τ) include:
R_x(0)≧R_x(τ).  (2)
R_x(τ)=R_S(τ)+R_N(τ)

[0027] If S(t) and N(t) are independent and both ergodic with zero mean, then for X(t)=S(t)+N(t):
R_x(τ)=R_S(τ)+R_N(τ)  (4)

[0028] The autocorrelation is for signal plus noise as represented in FIG. 4. Most random noise signals are not correlated, i.e., they satisfy: $\begin{array}{cc}\underset{r->\infty }{\lim }{R}_N\left(\tau \right)=0.& \left(5\right)\end{array}$
This is represented by autocorrelation in FIG. 5 as zero. Therefore, we have for large τ:
R_x(τ)≈R_s(τ)  (6)
Therefore, for large T, the noise has no correlation function. This property says that autocorrelation function has some noise immunity.

[0029] Frequency-Selective Autocorrelation Function

[0030] In real situation, direct application of autocorrelation function to utterance detector may not give enough robustness towards noises. The reasons include:

• • [0031] Many noise sources are not totally random. For instance, noises recorded in a moving car present some periodicity at low frequencies.
  • [0032] For computational reasons, the analysis window to implement autocorrelation is typically 30–50 ms, too short to attenuate low frequency noises. One solution to that is to pre-emphasize high frequency components. However, the pre-emphasis increases high frequency noise level.
  • [0033] Information leading to the determination of speech periodicity is mostly contained in a frequency band, corresponding to the range of human pitch period, rather than spread over the whole frequency range. However, this fact has not been used.

[0034] We apply a filter ƒ(τ) on the power spectrum of the autocorrelation function to attenuate the above-mentioned undesirable noisy components, as described by:
r_X(τ)=R_X(τ)*ƒ(τ)  (7)

[0035] To reduce the computation as in equation 1 and equation 7, the convolution is performed in the Discrete Fourier Transform (DFT) domain, as detailed below in the implementation. We can do the same by a DFT as illustrated in FIG. 6 by taking the signal and applying DFT, then do a frequency domain windowing following the equation 8 below and then do an inverse DFT to get the autocorrelation. The filter ƒ(τ) is specified in the frequency domain: $\begin{array}{cc}F\left(k\right)=\{\begin{array}{c}{\alpha }^{F_l-k}\mathrm{if}0\le k<F_l\\ 1\mathrm{if}{F}_l\le k<F_h\\ {\beta }^{k-F_h}\mathrm{if}{F}_h\le k<\frac{N}{2}\end{array}& \left(8\right)\end{array}$

• • [0036] with
    α=0.70  (9)
    β=0.85  (10)
    where F_l and F_h are respectively the discrete frequency indices under given sample frequency for 600 Hz and 1800 Hz.

[0037] We show two plots of r_X(τ) along with the time signal. The signal has been corrupted to 0 dB SNR. FIG. 7A shows a non-speech signal and FIG. 7B the frequency selective autocorrelation of the non-speech signal. FIG. 8A shows a speech signal and FIG. 8B the frequency selective autocorrelation function. It can be seen for the speech signal, a peak at 60 in FIG. 8B can be detected, with an amplitude substantially stronger than any peak in FIG. 7B.

[0038] Search for Periodicity

[0039] The periodicity measurement is defined as: $\begin{array}{cc}p=\underset{\tau =T_l}{\stackrel{T_h}{\max }}r\left(\tau \right)& \left(11\right)\end{array}$

[0040] T_l and T_h are pre-specified so that the period found will range from 75 Hz to 400 Hz. A larger value of p indicates a high energy level at the time index where p is found. We decide that the signal is speech if p is larger than a threshold.

[0041] The threshold is set to be 10 dB higher than a background noise level estimation:
θ=N+10  (12)

[0042] In FIG. 9, the curve “PRAM” shows the value of p for each of the incoming frames, the curve “DEC” shows the decision based on the threshold, and the curve “DIP” shows the evolution of the estimated background noise level.

[0043] Implementation

[0044] The calculation of the frame-wise decision is as follows:

• • [0045] 1. calculate the power spectrum of the signal
    • [0046] 1.1 filter the speech signal with H(z)=1−0.96_z⁻¹ (this filter is illustrated by FIG. 10).
    • [0047] 1.2 apply Hamming window $w\left(i\right)=0.54-0.46\cos \left(\frac{2\pi }{N}i\right)$
    • [0048] 1.3 perform FFT on the signal from step 1.2. X(k)=DPT(X(n) where X(k) has imaginary part Im and real part Re, k is the frequency index and n is time
    • [0049] 1.4 calculate the power spectrum which is |X(k)|²=Im²(X(k))+Re² (X(k))
  • [0050] 2. perform frequency shaping
    • [0051] 2.1 apply Eq-8 resulting R(k)
    • [0052] 2.2 $\forall k\in \left(0,\frac{N}{2}\right)R\left(\frac{N}{2}+k\right)=R\left(\frac{N}{2}-k\right)$
      to make R(k) symmetrical. As illustrated in FIG. 11 the third equation makes N/2 the center point. This is required to perform the inverse FFT.
  • [0053] 3. perform inverse FFT of R(k), resulting r_X() of Eq-7
  • [0054] 4. Search for p, the maximum of r_X() using Eq-11
  • [0055] 5. Calculate speech/non-speech decision S
    • [0056] 5.1 calculate the threshold using Eq-12
    • [0057] 5.2 (p>) decide “speech” else “non-speaker”.

[0058] Utterance-Level Detector 13 State-Machine

[0059] To make our final utterance detection, we need to incorporate some duration constraints about speech and non-speech. The two constants are used.

• • [0060] MIN-VOICE-SEG: the minimum number of frames to declare a speech segment.
  • [0061] MIN-PAUSE-SEG: the minimum number of frames to end a speech segment.

[0062] The functioning of the detector is completely described by a state machine. A state machine has a set of states connected by paths. Our state machine, shown in FIG. 12, has four states: non-speech; pre-speech, in-speech, and pre-nonspeech.

[0063] The machine has a current state, and based on the condition on the frame-wise speech/non-speech decision, will perform some action and move to a next state, as specified in Table 1.

[0064] In FIG. 12, the curve “STT: shows the state index, and the curve “LAB” labels the detected utterance.

[0065] In FIG. 12, one cycle means state. The arrow means to go to another state. The numbers are paths. Each path is defined by a condition. These are from level decisions. For each path, we need to take an action. Actions include state transitions. The action can be to do some calculation. After that action, we make a transition to the next state. In Table 1, the state is indicated by case. Suppose we need to make an utterance decision. We have four cases on states which are non-speech, pre-speech, in-speech and pre-nonspeech. We initialize on the left most case which is non-speech. We look at input. If the input frame is speech, we initialize a counter (n=1). In this case, we go to pre-speech state via path 2. If the frame level is non-speech, the system stays in the same state as represented by path 1. If in the pre-speech state and there is not enough counts of frames to indicate in-speech yet the frame is indicated or speech, we stay in pre-speech and increase the count by 1 as indicated by path 4. If the frame is speech and the count is N or greater (sufficiently long time), then it goes to the in-speech state as indicated by path 5. If the frame is not speech, then it takes the path 3 back to non-speech state. If we continue to detect speech at the frame level, we stay in the same state (path 6). If we receive a non-speech frame we move to pre-nonspeech state (path 7). If we again observe speech, we go back to in-speech state (path 8). If the next frame is non-speech, we stay in pre-nonspeech (path 9). If in pre-nonspeech for sufficiently long time (count of N) and frame input is below threshold, then we are in non-speech and the system goes to the non-speech state (path 10).

[0066] The utterance decision is represented by timing diagram (c) of FIG. 2.

[0067] We provide some pictures to show the difference between pre-emphasized energy and the proposed speech indicator based on frequency selective autocorrelation function.

[0068]

TABLE 1
case assignment and actions
CASE	CONDITION	ACTION	NEXT CASE	PATH
non-speech	S = speech	N = 1	Pre-speech	2
	Sγspeech	none	Non-speech	1
pre-speech	S = speech,	NpN + 1	Pre-speech	4
	N < MIN-VOICE-SEG
	S = speech, NμMIN-VOICE-SEG	start-extract	In-speech	5
	Sγspeech	none	Non-speech	3
in-speech	S = speech	none	In-speech	6
	Sγspeech	N = 1	Pre-non-speech	7
pre-nonspeech	S = speech	none	In-speech	8
	Sγspeech, N < MIN-PUASE-SEG	NpN + 1	Pre-non-speech	9
	Sγspeech, NμMIN-PAUSE-SEG	end-extract	Non-speech	10

[0069] FIG. 13 shows the time signal of an utterance with no noise added, 0 dB Signal to Noise Ratio (SNR) highway noise added, and 0 dB SNR white Gaussian noise added.

[0070] Basic Autocorrelation Function

[0071] FIG. 14 compares energy and the peak value obtained by directly searching Eq-1 for peak, i.e., using basic autocorrelation. It can be observed that basic autocorrelation function based on speech indicator gives significant lower background noise level, about 10, 15 and 15 dB lower for no noise added, highway noise added, and white Gaussian noise added, respectively. On the other hand, the difference for voiced speech is only a few dB.

[0072] For instance, for the highway noise case, the background noise level of energy contour is about 80 dB, and that of p is 65 dB. Therefore, p gives about 15 dB SNR improvement over energy.

[0073] Selective-Frequency Autocorrelation Function

[0074] FIG. 15 compares energy and the peak value obtained by Eq-11, i.e., using selective-frequency autocorrelation. It can be observed that improved autocorrelation function based speech indicator gives further lower background noise level, about 10, 35 and 20 dB lower for no noise added, highway noise added and white Gaussian noise added, respectively.

[0075] For instance, for the highway noise case, the background noise level of energy contour is about 80 dB, and that of p is 45 dB. Therefore, p gives about 35 dB SNR improvement over energy.

[0076] The difference of the two curves in each of the plots in FIG. 15 is plotted in FIG. 16. It can be seen that p gives consistent higher value than energy in voiced speech portion, especially in noisy situations.