Small array microphone for beam-forming and noise suppression

Abstract

Techniques are provided to suppress noise and interference using an array microphone and a combination of time-domain and frequency-domain signal processing. In one design, a noise suppression system includes an array microphone, at least one voice activity detector (VAD), a reference generator, a beam-former, and a multi-channel noise suppressor. The array microphone includes multiple microphones—at least one omni-directional microphone and at least one uni-directional microphone. Each microphone provides a respective received signal. The VAD provides at least one voice detection signal used to control the operation of the reference generator, beam-former, and noise suppressor. The reference generator provides a reference signal based on a first set of received signals and having desired voice signal suppressed. The beam-former provides a beam-formed signal based on a second set of received signals and having noise and interference suppressed. The noise suppressor further suppresses noise and interference in the beam-formed signal.

Description

BACKGROUND OF THE INVENTION

The present invention relates generally to communication, and more specifically to techniques for suppressing noise and interference in communication and voice recognition systems using an array microphone.

Communication and voice recognition systems are commonly used for many applications, such as hands-free car kit, cellular phone, hands-free voice control devices, telematics, teleconferencing system, and so on. These systems may be operated in noisy environments, such as in a vehicle or a restaurant. For each of these systems, one or multiple microphones in the system pick up the desired voice signal as well as noise and interference. The noise typically refers to local ambient noise. The interference may be from acoustic echo, reverberation, unwanted voice, and other artifacts.

Noise suppression is often required in many communication and voice recognition systems to suppress ambient noise and remove unwanted interference. For a communication or voice recognition system operating in a noisy environment, the microphone(s) in the system pick up the desired voice as well as noise. The noise is more severe for a hands-free system whereby the loudspeaker and microphone may be located some distance away from a talking user. The noise degrades communication quality and speech recognition rate if it is not dealt with in an appropriate manner.

For a system with a single microphone, noise suppression is conventionally achieved using a spectral subtract technique. For this technique, which performs signal processing in the frequency domain, the noise power spectrum of a noisy voice signal is estimated and subtracted from the power spectrum of the noisy voice signal to obtain an enhanced voice signal. The phase of the enhanced voice signal is set equal to the phase of the noisy voice signal. This technique is somewhat effective for stationary noise or slow-varying non-stationary (such as air-conditioner noise or fan noise, which does not change over time) but may not be effective for fast-varying non-stationary noise. Moreover, even for stationary noise, this technique can cause voice distortion if the noisy voice signal has a low signal-to-noise ratio (SNR). Conventional noise suppression for stationary noise is described in various literatures including U.S. Pat. Nos. 4,185,168 and 5,768,473.

For a system with multiple microphones, an array microphone is formed by placing these microphones at different positions sufficiently far apart. The array microphone forms a signal beam that is used to suppress noise and interference outside of the beam. Conventionally, the spacing between the microphones needs to be greater than a certain minimum distance D in order to form the desired beam. This spacing requirement prevents the array microphone from being used in many applications where space is limited. Moreover, conventional beam-forming with the array microphone is typically not effective at suppressing noise in an environment with diffused noise. Conventional systems with array microphone are described in various literatures including U.S. Pat. Nos. 5,371,789, 5,383,164, 5,465,302 and 6,002,776.

As can be seen, techniques that can effectively suppress noise and interference in communication and voice recognition systems are highly desirable.

SUMMARY OF THE INVENTION

Techniques are provided herein to suppress both stationary and non-stationary noise and interference using an array microphone and a combination of time-domain and frequency-domain signal processing. These techniques are also effective at suppressing diffuse noise, which cannot be handled by a single microphone system and a conventional array microphone system. The inventive techniques can provide good noise and interference suppression, high voice quality, and faster voice recognition rate, all of which are highly desirable for hands-free full-duplex applications in communication or voice recognition systems.

The array microphone is composed of a combination of omni-directional microphones and uni-directional microphones. The microphones may be placed close to each other (i.e., closer than the minimum distance required by a conventional array microphone). This allows the array microphone to be used in various applications. The array microphone forms a signal beam at a desired direction. This beam is then used to suppress stationary and non-stationary noise and interference.

A specific embodiment of the invention provides a noise suppression system that includes an array microphone, at least one voice activity detector (VAD), a reference generator, a beam-former, and a multi-channel noise suppressor. The array microphone is composed of multiple microphones, which include at least one omni-directional microphone and at least one uni-directional microphone. Each microphone provides a respective received signal. One of the received signals is designated as the main signal, and the remaining received signal(s) are designated as secondary signal(s). The VAD(s) provide at least one voice detection signal, which is used to control the operation of the reference generator, the beam-former, and the multi-channel noise suppressor. The reference generator provides a reference signal based on the main signal, a first set of at least one secondary signal, and an intermediate signal from the beam-former. The beam-former provides the intermediate signal and a beam-formed signal based on the main signal, a second set of at least one secondary signal, and the reference signal. Depending on the number of microphones used for the array microphone, the first and second sets may include the same or different secondary signals. The reference signal has the desired voice signal suppressed, and the beam-formed signal has the noise and interference suppressed. The multi-channel noise suppressor further suppresses noise and interference in the beam-formed signal to provide an output signal having much of the noise and interference suppressed.

In one embodiment, the array microphone is composed of three microphones—one omni-directional microphone and two uni-directional microphones (which may be placed close to each other). The omni-directional microphone is referred to as the main microphone/channel and its received signal is the main signal a(n). One of the uni-directional microphones faces toward a desired talker and is referred to as a first secondary microphone/channel. Its received signal is the first secondary signal s₁(n). The other uni-directional microphone faces away from the desired talker and is referred to as a second secondary microphone/channel. Its received signal is the second secondary signal s₂(n).

In another embodiment, the array microphone is composed of two microphones—one omni-directional microphone and one uni-directional microphone (which again may be placed close to each other). The uni-directional microphone faces toward the desired talker and its received signal is the main signal a(n). The omni-directional microphone is the secondary microphone/channel and its received signal is the secondary signal s(n).

Various other aspects, embodiments, and features of the invention are also provided, as described in further detail below.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows a diagram of a conventional array microphone system;

FIG. 2 shows a block diagram of a small array microphone system, in accordance with an embodiment of the invention;

FIGS. 3 and 4 show block diagrams of a first and a second voice activity detector;

FIG. 5 shows a block diagram of a reference generator and a beam-former;

FIG. 6 shows a block diagram of a third voice activity detector;

FIG. 7 shows a block diagram of a dual-channel noise suppressor;

FIG. 8 shows a block diagram of an adaptive filter;

FIG. 9 shows a block diagram of another embodiment of the small array microphone system; and

FIG. 10 shows a diagram of an implementation of the small array microphone system.

DESCRIPTION OF THE SPECIFIC EMBODIMENTS

For clarity, various signals and controls described herein are labeled with lower case and upper case symbols. Time-variant signals and controls are labeled with “(n)” and “(m)”, where n denotes sample time and m denotes frame index. A frame is composed of L samples. Frequency-variant signals and controls are labeled with “(k,m)”, where k denotes frequency bin. Lower case symbols (e.g., s(n) and d(m)) are used to denote time-domain signals, and upper case symbols (e.g., B(k,m)) are used to denote frequency-domain signals.

FIG. 1 shows a diagram of a conventional array microphone system 100. System 100 includes multiple (N) microphones 112a through 112n, which are placed at different positions. The spacing between microphones 112 is required to be at least a minimum distance of D for proper operation. A preferred value for D is half of the wavelength of the band of interest for the signal. Microphones 112a through 112n receive audio activity from a talking user 110 (which is often referred to as “near-end” voice or talk), local ambient noise, and unwanted interference. The N received signals from microphones 112a through 112n are amplified by N amplifiers (AMP) 114a through 114n, respectively. The N amplified signals are further digitized by N analog-to-digital converters (A/Ds or ADCs) 116a through 116n to provide N digitized signals s₁(n) through s_N(n).

The N received signals, provided by N microphones 112a through 112n placed at different positions, carry information for the differences in the microphone positions. The N digitized signals s₁(n) through S_N(n) are provided to a beam-former 118 and used to form a signal beam. This beam is used to suppress noise and interference outside of the beam and to enhance the desired voice within the beam. Beam-former 118 may be a fixed beam-former (e.g., a delay-and-sum beam-former) or an adaptive beam-former (e.g., an adaptive sidelobe cancellation beam-former). These various types of beam-former are well known in the art. Conventional array microphone system 100 is associated with several limitations that curtail its use and/or effectiveness, including (1) requirement of a minimum distance of D for the spacing between microphones and (2) marginal effectiveness for diffused noise.

FIG. 2 shows a block diagram of an embodiment of a small array microphone system 200. In general, a small array microphone system can include any number of microphones greater than one. Moreover, the microphones may be any combination of omni-directional microphones and uni-directional microphones. An omni-directional microphone picks up signal and noise from all directions. A uni-directional microphone picks up signal and noise from the direction pointed to by its main lobe. The microphones in system 200 may be placed closer than the minimum spacing distance D required by conventional array microphone system 100. For clarity, a small array microphone system with three microphones is specifically described below.

In the embodiment shown in FIG. 2, system 200 includes an array microphone that is composed of three microphones 212a, 212b, and 212c. More specifically, system 200 includes one omni-directional microphone 212b and two uni-directional microphones 212a and 212c. Omni-directional microphone 212b is referred to as the main microphone and is used to pick up desired voice signal as well as noise and interference. Uni-directional microphone 212a is the first secondary microphone which has its main lobe facing toward a desired talking user. Microphone 212a is used to pick up mainly the desired voice signal. Uni-directional microphone 212c is the second secondary microphone which has its main lobe facing away from the desired talker. Microphone 212c is used to pick up mainly the noise and interference.

Microphones 212a, 212b, and 212c provide three received signals, which are amplified by amplifiers 214a, 214b, and 214c, respectively. An ADC 216a receives and digitizes the amplified signal from amplifier 214a and provides a first secondary signal s₁(n). An ADC 216b receives and digitizes the amplified signal from amplifier 214b and provides a main signal a(n). An ADC 216c receives and digitizes the amplified signal from amplifier 214c and provides a second secondary signal s₂(n).

A first voice activity detector (VAD1) 220 receives the main signal a(n) and the first secondary signal s₁(n). VAD 1220 detects for the presence of near-end voice based on a metric of total power over noise power, as described below. VAD1220 provides a first voice detection signal d₁(n), which indicates whether or not near-end voice is detected.

A second voice activity detector (VAD2) 230 receives the main signal a(n) and the second secondary signal s₂(n). VAD2230 detects for the absence of near-end voice based on a metric of the cross-correlation between the main signal and the desired voice signal over the total power, as described below. VAD2230 provides a second voice detection signal d₂(n), which also indicates whether or not near-end voice is absent.

A reference generator 240 receives the main signal a(n), the first secondary signal s₁(n), the first voice detection signal d₁(n), and a first beam-formed signal b₁(n). Reference generator 240 updates its coefficients based on the first voice detection signal d₁(n), detects for the desired voice signal in the first secondary signal s₁(n) and the first beam-formed signal b₂(n), cancels the desired voice signal from the main signal a(n), and provides two reference signals r₁(n) and r₂(n). The reference signals r₁(n) and r₂(n) both contain mostly noise and interference. However, the reference signal r₂(n) is more accurate than r₁(n) in order to estimate the presence of noise and interference.

A beam-former 250 receives the main signal a(n), the second secondary signal s₂(n), the second reference signal r₂(n), and the second voice detection signal d₂(n). Beam-former 250 updates its coefficients based on the second voice detection signal d₂(n), detects for the noise and interference in the second secondary signal s₂(n) and the second reference signal r₂(n), cancels the noise and interference from the main signal a(n), and provides the two beam-formed signals b₁(n) and b₂(n). The beam-formed signal b₂(n) is more accurate than b₁(n) to represent the desired signal.

A delay unit 242 delays the second reference signal r₂(n) by a delay of T_a and provides a third reference signal r₃(n), which is r₃(n)=r₂(n−T_a). The delay T_a synchronizes (i.e., time-aligns) the third reference signal r₃(n) with the second beam-formed signal b₂(n).

A third voice activity detector (VAD3) 260 receives the third reference signal r₃(n) and the second beam-formed signal b₂(n). VAD3260 detects for the presence of near-end voice based on a metric of desired voice power over noise power, as described below. VAD3260 provides a third voice detection signal d₃(m) to dual-channel noise suppressor 280, which also indicates whether or not near-end voice is detected. The third voice detection signal d₃(m) is a function of frame index m instead of sample index n.

A dual-channel FFT unit 270 receives the second beam-formed signal b₂(n) and the third reference signal r₃(n). FFT unit 270 transforms the signal b₂(n) from the time domain to the frequency domain using an L-point FFT and provides a corresponding frequency-domain beam-formed signal B(k,m). FFT unit 270 also transforms the signal r₃(n) from the time domain to the frequency domain using the L-point FFT and provides a corresponding frequency-domain reference signal R(k,m).

A dual-channel noise suppressor 280 receives the frequency-domain signals B(k,m) and R(k,m) and the third voice detection signal d₃(m). Noise suppressor 280 further suppresses noise and interference in the signal B(k,m) and provides a frequency-domain output signal B_o(k,m) having much of the noise and interference suppressed.

An inverse FFT unit 290 receives the frequency-domain output signal B_o(k,m), transforms it from the frequency domain to the time domain using an L-point inverse FFT, and provides a corresponding time-domain output signal b_o(n). The output signal b_o(n) may be converted to an analog signal, amplified, filtered, and so on, and provided to a speaker.

FIG. 3 shows a block diagram of a voice activity detector (VAD1) 220x, which is a specific embodiment of VAD 1220 in FIG. 2. For this embodiment, VAD1220x detects for the presence of near-end voice based on (1) the total power of the main signal a(n), (2) the noise power obtained by subtracting the first secondary signal s₁(n) from the main signal a(n), and (3) the power ratio between the total power obtained in (1) and the noise power obtained in (2).

Within VAD 220x, a subtraction unit 310 subtracts the first secondary signal s₁(n) from the main signal a(n) and provides a first difference signal e₁(n), which is e₁(n)=a(n)−s₁(n). The first difference signal e₁(n) contains mostly noise and interference. High-pass filters 312 and 314 respectively receive the signals a(n) and e₁(n), filter these signals with the same set of filter coefficients to remove low frequency components, and provide filtered signals ã₁(n) and {tilde over (e)}₁(n), respectively. Power calculation units 316 and 318 then respectively receive the filtered signals ã₁(n) and {tilde over (e)}₁(n), compute the powers of the filtered signals, and provide computed powers p_a1(n) and p_e1(n), respectively. Power calculation units 316 and 318 may further average the computed powers. In this case, the averaged computed powers may be expressed as:p_a1(n)=a₁·p_a1(n−1)+(1−a₁)·ã₁(n)·ã₁(n), and  Eq (1a)p_e1(n)=a₁·p_e1(n−1)+(1−a₁)·{tilde over (e)}₁(n)·{tilde over (e)}₁(n),  Eq(1b)where α₁ is a constant that determines the amount of averaging and is selected such that 1>α₁>0. A large value for α₁ corresponds to more averaging and smoothing. The term p_a1(n) includes the total power from the desired voice signal as well as noise and interference. The term p_e1(n) includes mostly noise and interference power.

A divider unit 320 then receives the averaged powers p_a1(n) and p_e1(n) and calculates a ratio h₁(n) of these two powers. The ratio h₁(n) may be expressed as:

$\begin{array}{cc}{h}_1\left(n\right)=\frac{p_{\mathrm{a1}}\left(n\right)}{p_{\mathrm{e1}}\left(n\right)}.& \mathrm{Eq}\left(2\right)\end{array}$ The ratio h₁(n) indicates the amount of total power relative to the noise power. A large value for h₁(n) indicates that the total power is large relative to the noise power, which may be the case if near-end voice is present. A larger value for h₁(n) corresponds to higher confidence that near-end voice is present.

A smoothing filter 322 receives and filters or smoothes the ratio h₁(n) and provides a smoothed ratio h_s1(n). The smoothing may be expressed as:h_s1(n)=α_h1·h_s1(n−1)+(1−α_h1)·h₁(n),  Eq (3)where α_h1 is a constant that determines the amount of smoothing and is selected as 1>α_h1>0.

A threshold calculation unit 324 receives the instantaneous ratio h₁(n) and the smoothed ratio h_s1(n) and determines a threshold q₁(n). To obtain q₁(n), an initial threshold q₁′(n) is first computed as:

$\begin{array}{cc}{q}_1^{\prime }\left(n\right)=\{\begin{array}{cc}{\alpha }_{\mathrm{h1}}·q_1^{\prime }\left(n-1\right)+\left(1-{\alpha }_{\mathrm{h1}}\right)·h_1\left(n\right),& \mathrm{if}{h}_1\left(n\right)>{\beta }_1{h}_{\mathrm{s1}}\left(n\right)\\ q_1^{\prime }\left(n-1\right),& \mathrm{if}{h}_1\left(n\right)\underset{\_}{<}{\beta }_1{h}_{\mathrm{s1}}\left(n\right)\end{array},& \mathrm{Eq}\left(4\right)\end{array}$ where β₁ is a constant that is selected such that β₁>0. In equation (4), if the instantaneous ratio h₁(n) is greater than β₁h_s1(n), then the initial threshold q₁′(n) is computed based on the instantaneous ratio h₁(n) in the same manner as the smoothed ratio h_s1(n). Otherwise, the initial threshold for the prior sample period is retained (i.e., q₁′(n)=q₁′(n−1)) and the initial threshold q₁′(n) is not updated with h₁(n). This prevents the threshold from being updated under abnormal condition for small values of h₁(n).

The initial threshold q₁′(n) is further constrained to be within a range of values defined by Q_max1 and Q_min1. The threshold q₁(n) is then set equal to the constrained initial threshold q₁′(n), which may be expressed as:

$\begin{array}{cc}{q}_1\left(n\right)=\{\begin{array}{cc}{Q}_{\max 1},& \mathrm{if}{q}_1^{\prime }\left(n\right)>Q_{\max 1},\\ q_1^{\prime }\left(n\right),& \mathrm{if}{Q}_{\max 1}\underset{\_}{>}{q}_1^{\prime }\left(n\right)\underset{\_}{>}{Q}_{\min 1},\mathrm{and}\\ Q_{\min 1},& \mathrm{if}{Q}_{\min 1}>q_1^{\prime }\left(n\right),\end{array}& \mathrm{Eq}\left(5\right)\end{array}$ where Q_max1 and Q_min1 are constants selected such that Q_max1>Q_min1.

The threshold q₁(n) is thus computed based on a running average of the ratio h₁(n), where small values of h₁(n) are excluded from the averaging. Moreover, the threshold q₁(n) is further constrained to be within the range of values defined by Q_max1 and Q_min1. The threshold q₁(n) is thus adaptively computed based on the operating environment.

A comparator 326 receives the ratio h₁(n) and the threshold q₁(n), compares the two quantities h₁(n) and q₁(n), and provides the first voice detection signal d₁(n) based on the comparison results. The comparison may be expressed as:

$\begin{array}{cc}{d}_1\left(n\right)=\{\begin{array}{cc}1,& \mathrm{if}{h}_1\left(n\right)\ge q_1\left(n\right),\\ 0,& \mathrm{if}{h}_1\left(n\right)<q_1\left(n\right).\end{array}& \mathrm{Eq}\left(6\right)\end{array}$ The voice detection signal d₁(n) is set to 1 to indicate that near-end voice is detected and set to 0 to indicate that near-end voice is not detected.

FIG. 4 shows a block diagram of a voice activity detector (VAD2) 230x, which is a specific embodiment of VAD2230 in FIG. 2. For this embodiment, VAD2230x detects for the absence of near-end voice based on (1) the total power of the main signal a(n), (2) the cross-correlation between the main signal a(n) and the voice signal obtained by subtracting the main signal a(n) from the second secondary signal s₂(n), and (3) the ratio of the cross-correlation obtained in (2) over the total power obtained in (1).

Within VAD 230x, a subtraction unit 410 subtracts the main signal a(n) from the second secondary signal s₂(n) and provides a second difference signal e₂(n), which is e₂(n)=s₂(n)−a(n). High-pass filters 412 and 414 respectively receive the signals a(n) and e₂(n), filter these signals with the same set of filter coefficients to remove low frequency components, and provide filtered signals ã₂(n) and {tilde over (e)}₂(n), respectively. The filter coefficients used for high-pass filters 412 and 414 may be the same or different from the filter coefficients used for high-pass filters 312 and 314.

A power calculation unit 416 receives the filtered signal ã₂(n), computes the power of this filtered signal, and provides the computed power p_a2(n). A correlation calculation unit 418 receives the filtered signals ã₂(n) and {tilde over (e)}₂(n), computes their cross correlation, and provides the correlation p_ae(n). Units 416 and 418 may further average their computed results. In this case, the averaged computed power from unit 416 and the averaged correlation from unit 418 may be expressed as:p_a2(n)=α₂·p_a2(n−1)+(1−α₂)·ã₂(n)·ã₂(n), and  Eq (7a)p_ae(n)=α₂·p_ae(n−1)+(1−α₂)·ã₂(n)·{tilde over (e)}₂(n),  Eq (7b)where α₂ is a constant that is selected such that 1>α₂>0. The constant α₂ for VAD2230x may be the same or different from the constant α₁ for VAD1220x. The term p_a2(n) includes the total power for the desired voice signal as well as noise and interference. The term p_ae(n) includes the correlation between a(n) and e₂(n), which is typically negative if near-end voice is present.

A divider unit 420 then receives p_a2(n) and p_ae(n) and calculates a ratio h₂(n) of these two quantities, as follows:

$\begin{array}{cc}{h}_2\left(n\right)=\frac{p_{\mathrm{ae}}\left(n\right)}{p_{\mathrm{a2}}\left(n\right)}.& \mathrm{Eq}\left(8\right)\end{array}$

A smoothing filter 422 receives and filters the ratio h₂(n) to provide a smoothed ratio h_s2(n), which may be expressed as:h_s2(n)=α_h2·h_s2(n−1)+(1−α_h2)·h₂(n),  Eq(9)where α_h2 is a constant that is selected such that 1>α_h2>0. The constant α_h2 for VAD2230x may be the same or different from the constant α_h1 for VAD1220x.

A threshold calculation unit 424 receives the instantaneous ratio h₂(n) and the smoothed ratio h_s2(n) and determines a threshold q₂(n). To obtain q₂(n), an initial threshold q₂′(n) is first computed as:

$\begin{array}{cc}{q}_2^{\prime }\left(n\right)=\{\begin{array}{cc}{\alpha }_{\mathrm{h2}}·q_2^{\prime }\left(n-1\right)+\left(1+{\alpha }_{\mathrm{h2}}\right)·h_2\left(n\right),& \mathrm{if}{h}_2\left(n\right)>{\beta }_2{h}_{\mathrm{s2}}\left(n\right),\\ q_2^{\prime }\left(n-1\right),& \mathrm{if}{h}_2\left(n\right)\le {\beta }_2{h}_{\mathrm{s2}}\left(n\right),\end{array}& \mathrm{Eq}\left(10\right)\end{array}$ where β₂ is a constant that is selected such that β₂>0. The constant β₂ for VAD2230x may be the same or different from the constant β₁ for VAD 1220x. In equation (10), if the instantaneous ratio h₂(n) is greater than β₂h_s2(n), then the initial threshold q₂′(n) is computed based on the instantaneous ratio h₂(n) in the same manner as the smoothed ratio h_s2(n). Otherwise, the initial threshold for the prior sample period is retained.

The initial threshold q₂′(n) is further constrained to be within a range of values defined by Q_max2 and Q_min2. The threshold q₂(n) is then set equal to the constrained initial threshold q₂′(n), which may be expressed as:

$\begin{array}{cc}{q}_2\left(n\right)=\{\begin{array}{ccc}{Q}_{\max 2},& \mathrm{if}& q_2^{\prime }\left(n\right)>Q_{\max 2},\\ q_2^{\prime }\left(n\right),& \mathrm{if}{Q}_{\max 2}\ge & q_2^{\prime }\left(n\right)\ge Q_{\min 2},\mathrm{and}\\ Q_{\min 2,}& \mathrm{if}{Q}_{\min 2}>& q_2^{\prime }\left(n\right),\end{array}& \mathrm{Eq}\left(11\right)\end{array}$ where Q_max2 and Q_min2 are constants selected such that Q_max2>Q_min2.

A comparator 426 receives the ratio h₂(n) and the threshold q₂(n), compares the two quantities h₂(n) and q₂(n), and provides the second voice detection signal d₂(n) based on the comparison results. The comparison may be expressed as:

$\begin{array}{cc}{d}_2\left(n\right)=\{\begin{array}{cc}1,& \mathrm{if}{h}_2\left(n\right)\ge q_2\left(n\right),\\ 0,& \mathrm{if}{h}_2\left(n\right)<q_2\left(n\right).\end{array}& \mathrm{Eq}\left(12\right)\end{array}$ The voice detection signal d₂(n) is set to 1 to indicate that near-end voice is absent and set to 0 to indicate that near-end voice is present.

FIG. 5 shows a block diagram of a reference generator 240x and a beam-former 250x, which are specific embodiments of reference generator 240 and beam-former 250, respectively, in FIG. 2.

Within reference generator 240x, a delay unit 512 receives and delays the main signal a(n) by a delay of T₁ and provides a delayed signal a(n−T₁). The delay T₁ accounts for the processing delays of an adaptive filter 520. For linear FIR-type adaptive filter, T₁ is set to equal to half the filter length. Adaptive filter 520 receives the delayed signal a(n−T₁) at its x_in input, the first secondary signal s₁(n) at its x_ref input, and the first voice detection signal d₁(n) at its control input. Adaptive filter 520 updates its coefficients only when the first voice detection signal d₁(n) is 1. These coefficients are then used to isolate the desired voice component in the first secondary signal s₁(n). Adaptive filter 520 then cancels the desired voice component from the delayed signal a(n−T₁) and provides the first reference signal r₁(n) at its x_out output. The first reference signal r₁(n) contains mostly noise and interference. An exemplary design for adaptive filter 520 is described below.

A delay unit 522 receives and delays the first reference signal r₁(n) by a delay of T₂ and provides a delayed signal r₁(n−T₂). The delay T₂ accounts for the difference in the processing delays of adaptive filters 520 and 540 and the processing delay of an adaptive filter 530. Adaptive filter 530 receives the first beam-formed signal b₁(n) at its x_ref input, the delayed signal r₁(n−T₂) at its x_in input, and the first voice detection signal d₁(n) at its control input. Adaptive filter 530 updates its coefficients only when the first voice detection signal d₁(n) is 1. These coefficients are then used to isolate the desired voice component in the first beam-formed signal b₁(n). Adaptive filter 530 then further cancels the desired voice component from the delayed signal r₁(n−T₂) and provides the second reference signal r₂(n) at its x_out output. The second reference signal r₂(n) contains mostly noise and interference. The use of two adaptive filters 520 and 530 to generate the reference signals can provide improved performance.

Within beam-former 250x, a delay unit 532 receives and delays the main signal a(n) by a delay of T₃ and provides a delayed signal a(n−T₃). The delay T₃ accounts for the processing delays of adaptive filter 540. For linear FIR-type adaptive filter, T₃ is set to equal to half the filter length. Adaptive filter 540 receives the delayed signal a(n−T₃) at its x_in input, the second secondary signal s₂(n) at its x_ref input, and the second voice detection signal d₂(n) at its control input. Adaptive filter 540 updates its coefficients only when the second voice detection signal d₂(n) is 1. These coefficients are then used to isolate the noise and interference component in the second secondary signal s₂(n). Adaptive filter 540 then cancels the noise and interference component from the delayed signal a(n−T₃) and provides the first beam-formed signal b₁(n) at its x_out output. The first beam-formed signal b₁(n) contains mostly the desired voice signal.

A delay unit 542 receives and delays the first beam-formed signal b₁(n) by a delay of T₄ and provides a delayed signal b₁(n−T₄). The delay T₄ accounts for the total processing delays of adaptive filters 530 and 550. Adaptive filter 550 receives the delayed signal b₁(n−T₄) at its x_in input, the second reference signal r₂(n) at its x_ref input, and the second voice detection signal d₂(n) at its control input. Adaptive filter 550 updates its coefficients only when the second voice detection signal d₂(n) is 1. These coefficients are then used to isolate the noise and interference component in the second reference signal r₂(n). Adaptive filter 550 then cancels the noise and interference component from the delayed signal b₁(n−T₄) and provides the second beam-formed signal b₂(n) at its x_out output. The second beam-formed signal b₂(n) contains mostly the desired voice signal.

FIG. 6 shows a block diagram of a voice activity detector (VAD3) 260x, which is a specific embodiment of VAD3260 in FIG. 2. For this embodiment, VAD3260x detects for the presence of near-end voice based on (1) the desired voice power of the second beam-formed signals b₂(n) and (2) the noise power of the third reference signal r₃(n).

Within VAD 260x, high-pass filters 612 and 614 respectively receive the second beam-formed signal b₂(n) from beam-former 250 and the third reference signal r₃(n) from delay unit 242, filter these signals with the same set of filter coefficients to remove low frequency components, and provide filtered signals {tilde over (b)}₂(n) and {tilde over (r)}₃(n), respectively. Power calculation units 616 and 618 then respectively receive the filtered signals {tilde over (b)}₂(n) and {tilde over (r)}₃(n), compute the powers of the filtered signals, and provide computed powers p_b2(n) and p_r3(n), respectively. Power calculation units 616 and 618 may further average the computed powers. In this case, the averaged computed powers may be expressed as:p_b2(n)=α₃·p_b2(n−1)+(1−α₃)·{tilde over (b)}₂(n)·{tilde over (b)}₂(n), and  Eq(13a)p_r3(n)=α₃·p_r3(n−1)+(1−α₃)·{tilde over (r)}₃(n)·{tilde over (r)}₃(n),  Eq(13b)where α₃ is a constant that is selected such that 1>α₃>0. The constant α₃ for VAD3260x may be the same or different from the constant α₂ for VAD2230x and the constant α₁ for VAD1220x.

A divider unit 620 then receives the averaged powers p_b2(n) and p_r3(n) and calculates a ratio h₃(n) of these two powers, as follows:

$\begin{array}{cc}{h}_3\left(n\right)=\frac{p_{\mathrm{b2}}\left(n\right)}{p_{\mathrm{r3}}\left(n\right)}.& \mathrm{Eq}\left(14\right)\end{array}$ The ratio h₃(n) indicates the amount of desired voice power relative to the noise power.

A smoothing filter 622 receives and filters the ratio h₃(n) to provide a smoothed ratio h_s3(n), which may be expressed as:h_s3(n)=α_h3·h_s3(n−1)+(1−α_h3)·h₃(n),  Eq (15)where α_h3 is a constant that is selected such that 1>α_h3>0. The constant α_h3 for VAD3260x may be the same or different from the constant α_h2 for VAD2230x and the constant α_h1 for VAD1220x.

A threshold calculation unit 624 receives the instantaneous ratio h₃(n) and the smoothed ratio h_s3(n) and determines a threshold q₃(n). To obtain q₃(n), an initial threshold q₃′(n) is first computed as:

$\begin{array}{cc}{q}_3^{\prime }\left(n\right)=\{\begin{array}{cc}{\alpha }_{\mathrm{h3}}·q_3^{\prime }\left(n-1\right)+\left(1+{\alpha }_{\mathrm{h3}}\right)·h_3\left(n\right),& \mathrm{if}{h}_3\left(n\right)>{\beta }_3{h}_{\mathrm{s3}}\left(n\right),\\ q_3^{\prime }\left(n-1\right),& \mathrm{if}{h}_3\left(n\right)\le {\beta }_3{h}_{\mathrm{s3}}\left(n\right),\end{array}& \mathrm{Eq}\left(16\right)\end{array}$ where β₃ is a constant that is selected such that β₃>0. In equation (16), if the instantaneous ratio h₃(n) is greater than β₃h_s3(n), then the initial threshold q₃′(n) is computed based on the instantaneous ratio h₃(n) in the same manner as the smoothed ratio h_s3(n). Otherwise, the initial threshold for the prior sample period is retained.

The initial threshold q₃(n) is further constrained to be within a range of values defined by Q_max3 and Q_min3. The threshold q₃(n) is then set equal to the constrained initial threshold q₃′(n), which may be expressed as:

$\begin{array}{cc}{q}_3\left(n\right)=\{\begin{array}{ccc}{Q}_{\max 3},& \mathrm{if}& q_3^{\prime }\left(n\right)>Q_{\max 3},\\ q_3^{\prime }\left(n\right),& \mathrm{if}{Q}_{\max 3}\ge & q_3^{\prime }\left(n\right)\ge Q_{\min 3},\mathrm{and}\\ Q_{\min 3,}& \mathrm{if}{Q}_{\min 3}>& q_3^{\prime }\left(n\right).\end{array}& \mathrm{Eq}\left(17\right)\end{array}$ where Q_max3 and Q_min3 are constants selected such that Q_max3>Q_min3.

A comparator 626 receives the ratio h₃(n) and the threshold q₃(n) and averages these quantities over each frame m. For each frame, the ratio h₃(m) is obtained by accumulating L values for h₃(n) for that frame and dividing by L. The threshold q₃(m) is obtained in similar manner. Comparator 626 then compares the two averaged quantities h₃(m) and q₃(m) for each frame m and provides the third voice detection signal d₃(m) based on the comparison result. The comparison may be expressed as:

$\begin{array}{cc}{d}_3\left(m\right)=\{\begin{array}{cc}1,& \mathrm{if}{h}_3\left(m\right)\ge q_3\left(m\right),\\ 0,& \mathrm{if}{h}_3\left(m\right)<q_3\left(m\right).\end{array}& \mathrm{Eq}\left(18\right)\end{array}$ The third voice detection signal d₃(m) is set to 1 to indicate that near-end voice is detected and set to 0 to indicate that near-end voice is not detected. However, the metric used by VAD3 is different from the metrics used by VAD1 and VAD2.

FIG. 7 shows a block diagram of a dual-channel noise suppressor 280x, which is a specific embodiment of dual-channel noise suppressor 280 in FIG. 2. The operation of noise suppressor 280x is controlled by the third voice detection signal d₃(m).

Within noise suppressor 280x, a noise estimator 710 receives the frequency-domain beam-formed signal B(k,m) from FFT unit 270, estimates the magnitude of the noise in the signal B(k,m), and provides a frequency-domain noise signal N₁(k,m). The noise estimation may be performed using a minimum statistics based method or some other method, as is known in the art. The minimum statistics based method is described by R. Martin, in a paper entitled “Spectral subtraction based on minimum statistics,” EUSIPCO'94, pp. 1182–1185, September 1994. A noise estimator 720 receives the noise signal N₁(k,m), the frequency-domain reference signal R(k,m), and the third voice detection signal d₃(m). Noise estimator 720 determines a final estimate of the noise in the signal B(k,m) and provides a final noise estimate N₂(k,m), which may be expressed as:

$\begin{array}{cc}{N}_2\left(k,m\right)=\{\begin{array}{cc}{\gamma }_{\mathrm{a1}}·N_1\left(k,m\right)+{\gamma }_{\mathrm{a2}}·R\left(k,m\right),& \mathrm{if}{d}_3\left(m\right)=1,\\ {\gamma }_{\mathrm{b1}}·N_1\left(k,m\right)+{\gamma }_{\mathrm{b2}}·R\left(k,m\right),& \mathrm{if}{d}_3\left(m\right)=0,\end{array}& \mathrm{Eq}\left(19\right)\end{array}$ where γ_a1, γ_a2, γ_b1, and γ_b2 are constants and are selected such that γ_a1>γ_b1>0 and γ_b2>γ_a2>0. As shown in equation (19), the final noise estimate N₂(k,m) is set equal to the sum of a first scaled noise estimate, γ_x1·N₁(k,m), and a second scaled noise estimate, γ_x2·|R(k,m)|, where γ_x can be equal to γ_a or γ_b. The constants γ_a1, γ_a2, γ_b1, and γ_b2 are selected such that the final noise estimate N₂(k,m) includes more of the noise estimate N₁(k,m) and less of the reference signal magnitude |R(k,m)| when d₃(m)=1, indicating that near-end voice is detected. Conversely, the final noise estimate N₂(k,m) includes less of the noise estimate N₁(k,m) and more of the reference signal magnitude |R(k,m)| when d₃(m)=0, indicating that near-end voice is not detected.

A noise suppression gain computation unit 730 receives the frequency-domain beam-formed signal B(k,m), the final noise estimate N₂(k,m), and the frequency-domain output signal B_o(k, m−1) for a prior frame from a delay unit 734. Computation unit 730 computes a noise suppression gain G(k,m) that is used to suppress additional noise and interference in the signal B(k,m).

To obtain the gain G(k,m), an SNR estimate G′_SNR,B(k,m) for the beam-formed signal B(k,m) is first computed as follows:

$\begin{array}{cc}{G}_{\mathrm{SNR},B}^{\prime }\left(k,m\right)=\frac{B\left(k,m\right)}{N_2\left(k,m\right)}-1.& \mathrm{Eq}\left(20\right)\end{array}$ The SNR estimate G′_SNR,B(k,m) is then constrained to be a positive value or zero, as follows:

$\begin{array}{cc}{G}_{\mathrm{SNR},B}\left(k,m\right)=\{\begin{array}{cc}{G}_{\mathrm{SNR},B}^{\prime }\left(k,m\right),& \mathrm{if}{G}_{\mathrm{SNR},B}^{\prime }\left(k,m\right)\ge 0,\\ 0,& \mathrm{if}{G}_{\mathrm{SNR},B}^{\prime }\left(k,m\right)<0.\end{array}& \mathrm{Eq}\left(21\right)\end{array}$

A final SNR estimate G_SNR(k,m) is then computed as follows:

$\begin{array}{cc}{G}_{\mathrm{SNR}}\left(k,m\right)=\frac{\lambda ·B_o\left(k,m-1\right)}{N_2\left(k,m\right)}+\left(1-\lambda \right)·G_{\mathrm{SNR},B}\left(k,m\right),& \mathrm{Eq}\left(22\right)\end{array}$ where λ is a positive constant that is selected such that 1>λ>0. As shown in equation (22), the final SNR estimate G_SNR(k,m) includes two components. The first component is a scaled version of an SNR estimate for the output signal in the prior frame, i.e., λ·|B_o(k, m−1)|/N₂(k,m). The second component is a scaled version of the constrained SNR estimate for the beam-formed signal, i.e., (1−λ)·G_SNR,B(k,m). The constant λ determines the weighting for the two components that make up the final SNR estimate G_SNR(k,m).

The gain G(k,m) is then computed as:

$\begin{array}{cc}G\left(k,m\right)=\frac{G_{\mathrm{SNR}}\left(k,m\right)}{1+G_{\mathrm{SNR}}\left(k,m\right)}.& \mathrm{Eq}\left(23\right)\end{array}$ The gain G(k,m) is a real value and its magnitude is indicative of the amount of noise suppression to be performed. In particular, G(k,m) is a small value for more noise suppression and a large value for less noise suppression.

A multiplier 732 then multiples the frequency-domain beam-formed signal B(k,m) with the gain G(k,m) to provide the frequency-domain output signal B_o(k,m), which may be expressed as:B_o(k,m)=B(k,m)·G(k,m)  Eq (24)

FIG. 8 shows a block diagram of an embodiment of an adaptive filter 800, which may be used for each of adaptive filters 520, 530, 540, and 550 in FIG. 5. Adaptive filter 800 includes a FIR filter 810, summer 818, and a coefficient computation unit 820. An infinite impulse response (IIR) filter or some other filter structure may also be used in place of the FIR filter. In FIG. 8, the signal received on the x_ref input is denoted as x_ref(n), the signal received on the x_in input is denoted as x_in(n), the signal received on the control input is denoted as d(n), and the signal provided to the x_out, output is denoted as x_out(n).

Within FIR filter 810, the digital samples for the reference signal x_ref(n) are provided to M−1 series-coupled delay elements 812b through 812m, where M is the number of taps of the FIR filter. Each delay element provides one sample period of delay. The reference signal x_ref(n) and the outputs of delay elements 812b through 812m are provided to multipliers 814a through 814m, respectively. Each multiplier 814 also receives a respective filter coefficient h_i(n) from coefficient calculation unit 820, multiplies its received samples with its filter coefficient h_i(n), and provides output samples to a summer 816. For each sample period n, summer 816 sums the M output samples from multipliers 814a through 814m and provides a filtered sample for that sample period. The filtered sample x_fir(n) for sample period n may be computed as:

$\begin{array}{cc}{x}_{\mathrm{fir}}\left(n\right)=\sum _{i=0}^{M-1}{h}_i^{*}·x_{\mathrm{ref}}\left(n-i\right),& \mathrm{Eq}\left(25\right)\end{array}$ where the symbol “*” denotes a complex conjugate. Summer 818 receives and subtracts the FIR signal x_fir(n) from the input signal x_in(n) and provides the output signal x_out(n).

Coefficient calculation unit 820 provides the set of M coefficients for FIR filter 810, which is denoted as H*(n)=[h₀*(n), h₁*(n), . . . h_M−1*(n)]. Unit 820 further updates these coefficients based on a particular adaptive algorithm, which may be a least mean square (LMS) algorithm, a normalized least mean square (NLMS) algorithm, a recursive least square (RLS) algorithm, a direct matrix inversion (DMI) algorithm, or some other algorithm. The NLMS and other algorithms are described by B. Widrow and S. D. Sterns in a book entitled “Adaptive Signal Processing,” Prentice-Hall Inc., Englewood Cliffs, N.J., 1986. The LMS, NLMS, RLS, DMI, and other adaptive algorithms are described by Simon Haykin in a book entitled “Adaptive Filter Theory”, 3rd edition, Prentice Hall, 1996. Coefficient update unit 820 also receives the control signal d(n) from VAD1 or VAD2, which controls the manner in which the filter coefficients are updated. For example, the filter coefficients may be updated only when voice activity is detected (i.e., when d(n)=1) and may be maintained when voice activity is not detected (i.e., when d(n)=0).

For clarity, a specific design for the small array microphone system has been described above, as shown in FIG. 2. Various alternative designs may also be provided for the small array microphone system, and this is within the scope of the invention. These alternative designs may include fewer, different, and/or additional processing units than those shown in FIG. 2. Also for clarity, specific embodiments of various processing units within small array microphone system 200 have been described above. Other designs may also be used for each of the processing units shown in FIG. 2, and this is within the scope of the invention. For example, VAD1 and VAD3 may detect for the presence of near-end voice based on some other metrics than those described above. As another example, reference generator 240 and beam-former 250 may be implemented with different number of adaptive filters and/or different designs than the ones shown in FIG. 5.

FIG. 9 shows a diagram of an embodiment of another small array microphone system 900. System 900 includes an array microphone composed of two microphones 912a and 912b. More specifically, system 900 includes one omni-directional microphone 912a and one uni-directional microphone 912b, which may be placed close to each other (i.e., closer than the distance D required for the conventional array microphone). Uni-directional microphone 912b is the main microphone which has a main lobe facing toward the desired talker. Microphone 912b is used to pick up the desired voice signal. Omni-directional microphone 912a is the secondary microphone. Microphones 912a and 912b provide two received signals, which are amplified by amplifiers 914a and 914b, respectively. An ADC 916a receives and digitizes the amplified signal from amplifier 914a and provides the secondary signal s₁(n). An ADC 916b receives and digitizes the amplified signal from amplifier 914b and provides the main signal a(n). The noise and interference suppression for system 900 may be performed as described in the aforementioned U.S. patent application Ser. No. 10/371,150.

FIG. 10 shows a diagram of an implementation of a small array microphone system 1000. In this implementation, system 1000 includes three microphones 1012a through 1012c, an analog processing unit 1020, a digital signal processor (DSP) 1030, and a memory 1032. Microphones 1012a through 1012c may correspond to microphones 212a through 212c in FIG. 2. Analog processing unit 1020 performs analog processing and may include amplifiers 214a through 214c and ADCs 216a through 216c in FIG. 2. Digital signal processor 1030 may implement various processing units used for noise and interference suppression, such as VAD1220, VAD2230, VAD3260, reference generator 240, beam-former 250, FFT unit 270, noise suppressor 280, and inverse FFT unit 290 in FIG. 2. Memory 1032 provides storage for program codes and data used by digital signal processor 1030.

The array microphone and noise suppression techniques described herein may be implemented by various means. For example, these techniques may be implemented in hardware, software, or a combination thereof. For a hardware implementation, the processing units used to implement the array microphone and noise suppression may be implemented within one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processors, controllers, micro-controllers, microprocessors, other electronic units designed to perform the functions described herein, or a combination thereof.

For a software implementation, the array microphone and noise suppression techniques may be implemented with modules (e.g., procedures, functions, and so on) that perform the functions described herein. The software codes may be stored in a memory unit (e.g., memory unit 1032 in FIG. 10) and executed by a processor (e.g., DSP 1030).

The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A noise suppression system comprising: an array microphone comprised of a plurality of microphones and operative to provide a plurality of received signals, one received signal for each microphone, wherein the plurality of microphones include at least one omni-directional microphone and at least one uni-directional microphone;at least one voice activity detector operative to provide first and second voice detection signals based on the plurality of received signals;a reference generator operative to provide a reference signal based on the first voice detection signal and a first set of received signals selected from among the plurality of received signals;a beam-former operative to provide a beam-formed signal based on the second voice detection signal, the reference signal, and a second set of received signals selected from among the plurality of received signals, wherein the beam-formed signal has noise and interference suppressed; anda multi-channel noise suppressor operative to further suppress noise and interference in the beam-formed signal and provide an output signal.

2. The system of claim 1, wherein the reference generator is operative to provide the reference signal having substantially noise and interference, and wherein the beam-former is operative to suppress the noise and interference in the beam-formed signal using the reference signal.

3. The system of claim 1, wherein the reference generator includes a first set of at least one adaptive filter operative to filter the first set of received signals and an intermediate signal from the beam-former to provide the reference signal, and wherein the beam-former includes a second set of at least one adaptive filter operative to filter the second set of received signals and the reference signal to provide the beam-formed signal.

4. The system of claim 1, wherein the reference generator and the beam-former are operative to perform time-domain signal processing.

5. The system of claim 1, wherein the multi-channel noise suppressor is operative to perform frequency-domain signal processing.

6. The system of claim 1, wherein the multi-channel noise suppressor is operative to derive a gain value indicative of an estimated amount of noise and interference in the beam-formed signal and to suppress the noise and interference in the beam-formed signal with the gain value.

7. The system of claim 1, wherein the estimated amount of noise and interference in the beam-formed signal is determined based on the reference signal, the beam-formed signal, and the output signal.

8. The system of claim 1, wherein the at least one voice activity detector includes a first voice activity detector operative to provide the first voice detection signal based on the first set of received signals.

9. The system of claim 8, wherein the first voice detection signal is determined based on a ratio of total power over noise power.

10. The system of claim 8, wherein the at least one voice activity detector further includes a second voice activity detector operative to provide the second voice detection signal based on the second set of received signals.

11. The system of claim 10, wherein the second voice detection signal is determined based on a ratio of cross-correlation between a desired signal and a main signal over total power.

12. The system of claim 8, wherein the at least one voice activity detector further includes a third voice activity detector operative to provide a third voice detection signal based on the reference signal and the beam-formed signal, and wherein the multi-channel noise suppressor is operative to suppress noise and interference in the beam-formed signal based on the third voice detection signal.

13. The system of claim 12, wherein the third voice detection signal is determined based on a power ratio of the beam-formed signal over a reference noise signal.

14. The system of claim 1, wherein the array microphone comprises one omni-directional microphone and two uni-directional microphones.

15. The system of claim 14, wherein the omni-directional microphone is designated as a main channel and the two unidirectional microphones are designated as secondary channels.

16. The system of claim 14, wherein one of the two unidirectional microphones faces toward a voice signal source and the other one of the two uni-directional microphones faces away from the voice signal source.

17. The system of claim 16, wherein the first set of received signals includes a main received signal from the omni-directional microphone and a first secondary received signal from the uni-directional microphone facing toward the voice signal source, and wherein the second set of received signals includes the main received signal and a second secondary received signal from the uni-directional microphone facing away from the voice signal source.

18. The system of claim 1, wherein the array microphone comprises one omni-directional microphone and one uni-directional microphone.

19. The system of claim 18, wherein the uni-directional microphone faces toward a voice signal source, and wherein the first and second sets of received signals both include a main received signal from the uni-directional microphone and a secondary received signal from the omni-directional microphone.

20. An apparatus comprising: means for obtaining a plurality of received signals from a plurality of microphones forming an array microphone, wherein the plurality of microphones include at least one omni-directional microphone and at least one uni-directional microphone;means for providing first and second voice detection signals based on the plurality of received signals;means for providing a reference signal based on the first voice detection signal and a first set of received signals selected from among the plurality of received signals;means for providing a beam-formed signal based on the second voice detection signal, the reference signal, and a second set of received signals selected from among the plurality of received signals, wherein the beam-formed signal has noise and interference suppressed; andmeans for suppressing additional noise and interference in the beam-formed signal to provide an output signal.

21. The apparatus of claim 20, wherein the plurality of microphones include one omni-directional microphone and two uni-directional microphones, and wherein one of the two uni-directional microphones faces toward a voice signal source and the other one of the two uni-directional microphones faces away from the voice signal source.

22. A method of suppressing noise and interference, comprising: obtaining a plurality of received signals from a plurality of microphones forming an array microphone, wherein the plurality of microphones include at least one omni-directional microphone and at least one uni-directional microphone;providing first and second voice detection signals based on the plurality of received signals;providing a reference signal based on the first voice detection signal and a first set of received signals selected from among the plurality of received signals;providing a beam-formed signal based on the second voice detection signal, the reference signal, and a second set of received signals selected from among the plurality of received signals, wherein the beam-formed signal has noise and interference suppressed; andsuppressing additional noise and interference in the beam-formed signal to provide an output signal.

23. The method of claim 22, wherein the reference signal and beam-formed signal are provided using time-domain signal processing, and wherein the suppressing is performed using frequency-domain signal processing.