Acoustic Signal Processing - Science topic

Isam Alkhalifawi Sir, these adaptive echo cancellation techniques don't help me in my scenario. My echo is non-stationary and the same sound is being produced with delay by multiple speakers, not just 1 speaker.

Please, try https://arxiv.org/ftp/arxiv/papers/1709/1709.07552.pdf.

May be, it will be usefull...

Avisoft SasLab Pro 5.2 software has got inbuilt lowpass, highpass, notchpass and bandpass filters critical in elimination of unwanted noise. Open the Avisoft SasLab Pro 5.2 software, the choose the edit menu. Under edit menu select Filter. Under filter, choose the Time domain IIR or FIR Filter. If its the background noise, then you may record the room tone and filter it out. Similarly the Avisoft UltraSound Gate allows you to attenuate unwanted signals.

Dear Sheng,

welcome,

I think you had already solved the problem. However i would like to introduce a solution. There are two types of electrical vibration harvesters. The electrostaic MEMES converter and the piezoelectric transducer. You can use four flat converters to collect the sound from all sides. You can also put the sound line in in the focus of parabolic reflector and receiver the reflected sound by one flat converters. The second solution may be better. as it uses only one flat converter.

Best wishes

Thanks for your kind response, Silber. I have searched for the corpus links but i couldn't find the link. Can you share the url link related with free corpus maintained by carnegie mellon university that will be helpful.

When you describe what are your streams and why you want to synchronize them i can make you a proposal. However, when you work in packet switched network

Then you can follow the synchronization techniques used in it. It is so that the packets are sealed and addressed and numbered such that one can assemble them at the destination.

Best wishes

You want to increase the size of the audio frame to contain the whole audio message. At first i think such question needs investigation to answer it.

But i think the frame size is determined by the latency in the transmission system.

It is required to carry out all the signal processing in such latency time. One other important point is that the transmission medium may be dynamic such that the decoding may be very complicated or even impossible if we increased the frame size.. I think the size of the frames are dictated more by the more complex decoding process that the coding process. Also, the transmission medium imposes restriction on the frame size..

Also the cost of the processing even if it increases only proportional to the size of the frame must be considered as one can not use powerful computing platforms for all communication equipment.

An another important point that audio signal is not a continuous signal in nature but contains interruptions.

I think the optimum frame size varies among applications.

Best wishes

Dear Nesrin,

This problem can be solved using the numerical method previously used in CAE IMPULSE:

This approach has good stability and very low computational costs (time and memory).

The simulation time on standard desktop computers is about 1-10 minutes (depending on the specific task).

Regards

Please share me the best answer might you get...

It sounds simple, but unfortunately, it is not!

There are many confounding factors that make this process complicated. I give you some examples: consider you have a recording of your own voice recorded in a sound proof room saying "OPEN THE DOOR", and you would like to use that recording as the reference to which other voice commands are compared to take an action to open the door, for example.

Considering the formants of the two signals and comparing them using some similarity measures could be a very simple and quick solution. But unfortunately, they do not provide a good result since, for example, the similarity measure of two completely different sentences recorded in a particular acoustic environment can be relatively higher than two roughly similar sentences recorded in different environment, or if in the second recording the speaker utters the similar words than the reference recording but in different order.

To deal with these factors and variabilities, you might need a model (such as hidden Markov model, Gaussian mixture model) to capture the acoustical characteristics of the signals (in some relevant feature space such as cepstral domain or time-frequency domain) and to relate the segments of a signal to the language unites, and also you need a language model to link the unites to recognize the sentence. All these procedures are covered under the speech recognition field.

Some how you don't get my reply

Ori Yeheskel

There is also this: http://www.oxfordwaveresearch.com

Just copy an paste data.

Create a variable in the workspace then paste the excel data into the variable editor window (don't paste the header).

You can open the excel file from matlab and import the data too, each column will become a variable.

After that you can apply the corrcoef function.

Dear Haidar,

You may use MATLAB and apply the two microphone based sound intensity calculation procedure for processing your data. You could find some useful resources for this purpose at Bruel&Kjaer website. In addition, for your short reference, please read the frequency domain Kalman filter as proposed by Bai et al (J. Ac. Soc. Am. 133(3), March 2013 pp 1425-1432. 

Good Morning, Prashanth

As your question is vague, I would assume that you are not referring to removal of noise in acoustic sound imagery but to pure acoustic signals.

 Here is a patent that has been granted recently and it can help you start thinking of how you can apply the knowledge taken from  the patent.

 My understanding is that you want to apply two-dimensional filtering to stereo sound ( or 3D filters to 3D sound).

 2D filters can be applied as separable or non-separable. You can apply separable filters in each one of the dimensions and the non-separable 2d filters similar to the application in imaging. Sometimes you can apply their inverse form and use FFT.

Here is an example in an article from Ieee explore:

Good luck, please do not hesitate to ask for more, Viara

Hi Milan,

What is your sound reproduction system? When do you say 60 dB is SPL?

If your reproduction system is headphones, you can use a binaural head to measure the sound pressure level at each ear and calibrate your system. In the case of loudspeakers, I would measure the SPL at the listener position using a sound level meter.

Best regards,

Diego

Please try https://www.mathworks.com/matlabcentral/fileexchange/47333-pesq-matlab-driver?requestedDomain=www.mathworks.com or https://www.mathworks.com/matlabcentral/fileexchange/33820-pesq-matlab-wrapper?focused=5203883&tab=function

Hi

The concept of Flame Transfer Function is new to me. I just found this paper and had a quick look at it where the speaker is used to modulate the flow through the nozzle. 

It seems to me that you might be able to detect part of the information you want on the downstream side using reciprocity but, I may be off so treat the above with a grain of salt. 

/C

You can detect point on crack initiation and hence initiation toughness but you can't generate complete j-r curve.

No.  This is wrong.  The step is still 15 MHz, so the total time is still 66.66666667 ns.  Because you have more points (and thus the bandwidth is bigger) the time step is now 248.8 ps.

step  = 15 MHz

start - stop = 4.005 GHz

4.005 GHz / 15 MHz = 267 steps

267 steps +1 = 268 points

total range of time = 1/15 MHz = 66.666666666667 ns

time step = 66.666666667 ns / 268 = 248.8 ps = 1/4.005 GHz *267/268

The zero padding won't have changed the shape of the time response in proportion to the total time, because no extra data has been added or taken away (it will change it a little bit if you used a non-constant window function that is over the whole frequency range rather than just over where the data is), but the time sample points will be at a different place on the curve.  Zero padding puts the time points closer together.  It is used to make the curves look smoother.  Zero padding to 4 or 8 times the length is sometimes used.

Putting the two columns beside each other makes me wonder if you think the times and frequencies correspond with each other.  No particular frequency corresponds to any particular time.  Every frequency is used to calculate the result at each time.  In fact each time data point is just basically the sum of the signals at all the frequencies (with the measured amplitudes and phases) at that time.

To my knowledge a SNR > 30 dB is considered as a clean speech. Some guidelines can be found in the link.

What you actually want to change is the window size. For that you need to change the pitch settings and give a minimum pitch 4 times more than the rate you want rate you want to have (i.e. 400/4000 for a window size of 0.ü1/0.001). This will however not change the overall mean intensity much, and is more relevant to clinical research.

The exponential functions are useful not only in the study of acoustical signals but in the study of "linear" systems which are generally considered because these function are not modified by a linear system: if the input of a linear system is an exponential function the output is the same exponential function multiplied by a complex number whose modulus give the "gain" of the system and the phases the phase change due to the system...

If the recorded speech signal is degraded by Background noise, using wavelet tresholding would be more appropriate than low pass filters for vowels.

Hi Dr Victor,

SAW in my case is not generated using an IDT , instead it is optically generated. I want to use IDT to detect such an optically induced SAW.

Thanks

Good question and good answer :)

Perhaps a little late, but there is a toolbox available which contains matlab implementations of a number of loudness models, which I've found useful in the past. http://genesis-acoustics.com/en/loudness_online-32.html

Best way I know of is what Sina said with the CPSD or the coherence. To do this with MATLAB's built in functions, use the "cpsd" for the cross-power spectral density or mscohere function for the coherence.

If they're noisy signals as you say, you'll may want to use a Hanning window with some overlap for the window edge effects. We usually use 75% overlap for aeroacoustic broadband signals.

As you focus on sound analysis, you may find interesting nowadays state-of-the-art technique for improving speech recognition acoustic models by augmenting training data using with speed, frequency and tempo warping, which is indirectly related to synthetic data generation ( http://www.danielpovey.com/files/2015_interspeech_augmentation.pdf )

Another example of creating synthetic data is block mixing that seems to be useful in polyphonic acoustic event detection: https://arxiv.org/pdf/1604.00861.pdf

We also tried the latter in recent DCASE challenge gaining a little bit on F-measure, but not as large as the author of previous work: https://www.researchgate.net/publication/306118891_DCASE_2016_Sound_Event_Detection_System_Based_on_Convolutional_Neural_Network

There are two excellent alternatives (free software):

For R users, the package "seewave": http://rug.mnhn.fr/seewave/

The Brazilian biologist Marcos Gridi Papp developed Sound Ruler: http://soundruler.sourceforge.net/main/

Both are excellent options.

Of course there are commercial software (most of them with free "light" versions)

Best

Hi

Perhaps you can find something useful in the below links where I believe they have been barking up the same tree?

Sincerely

Claes

Yes, if you need to.

BB-C 

Dear Fernando,

Attachments (with calculations info) are available. 

MATLAB code for calculations is relatively simple and is given below.

[Pd,Pfa] = rocsnr(30);

idx = find(Pfa==0.005); % find index for Pfa=0.005

sprintf('%.9f',Pd(idx))

The result for calculation looks as follows (I expected to get 0.998).

ans =

Empty matrix: 1-by-0

After getting this result I tried to increase Pfa value, but the result is 1.

[Pd,Pfa] = rocsnr(30);

idx = find(Pfa==0.01); % find index for Pfa=0.01

sprintf('%.9f',Pd(idx))

ans =

1.000000000

The links to MATLAB documentation related to ROC curves are given below.

It is interesting that ROC curves were first introduced in MATLAB R2011a.

Sincerely,

Oleksandr

I think you will find that you are not seeing the actual string vibrations but strobe effects caused by the sample rate of the digital camera.  The strings are not bending in that rapid pattern, but move from side to side in the time it takes the camera to scan a small distance up the string (actually across its own image plane).  Similarly there are not waves moving along the string in the way it looks.  These are all to do with the sampling in the camera - just like wheels turn backwards in some films.  The A string has a frequency of 110 Hz.  The camera only looks at it 30 times a second, it does not quite four vibrations between each frame.  It will be in sync with the camera at different points in the picture for each frame, that's why you see the pattern travel along the string.  The camera lies - especially digital cameras.

Hello Mr. Alejandro. My diffusers is not surface circular but surface square and surface area 60 cm x 60 cm according to the direction Vorländer and Mommertz in paper "Defnition and measurement of random-incidence scattering coecients". Thanks for posting the email address, maybe we are communication via email.

Best regards. Sofian hanafi

Dear Gaurav,

Low pass filtering may help a little (as described above).

However, if I understand your experiment correctly, then you want to measure impact. i.e. a fast acceleration and deceleration. Low pass filters can attenuate the peak signal and so give a confusing answer.

A better solution is to identify the source of noise and tackle that. the following check list may help.

1). is the noise mains frequency? if so use better shielding at the transducer and over the leads to the amplifier

2). is the noise pick up from other machines in the same lab; use better shielding place the amplifier as close to the transducer as possible, with screened cable. Do not use the screen as part of the measurement circuit. Use differential amplifier inputs wherever possible.

3). can you increase the signal from the transducer? (this depends on the type of Tr) on some you can increase the excitation voltage. (but be careful of dissipation, one can just increase voltage around the impact incident)

4). Are you using the right transducer? are there any trade-offs you can take advantage of?

5). do a FFT (Fast Fourier transform) of the noise to see if there are any dominant frequencies you can filter out using a notch filter

6). Is the noise coming from the A to D converter (typically clock frequency noise), if so check your screening and grounding, use an analogue amplifier close to the transducer to increase signal before the A to D.

One last comment, its easier to get a clean signal from the transducer. Trying to filter a noisy signal is unlikely to give you the best result.

By all means come back to me if you have more detail of your set up and let me know what measurement accuracy you are looking for.

Regards

Paul

Hi David

Debugging software always is dog's work.

Suggestions on simple things to compare are

Where possible, try and swap data between systems, i.e. get test data to act as simulation data and vice versa. If this works, use a simple calibrator signal of 1 g at fixed and known frequency as first signal to test AMESim.

If I should venture a guess, test systems use fixed sampling rates as the measurement HW uses fixed clock frequency. A simulation system may use different time steps because of numerical efficiency and this difference in time stepping may complicate post processing.

Have fun

Claes

A priori, there is no correlation between the harmonics and the formant frequencies of a spoken word. Imagine you are saying a vowel (such as /a/) on different pitches. The formant frequencies remain more of less constant while the fundamental frequencies f0 and upper harmonics vary (if f0 is in the speech range). The coincidence of 2*f0 and F2 can happen but it can not be a general case. In the singing voice, the belting technique (such as in Bulgarian singing or in music theatre singing style) is characterized by the tuning of the *first* formant (F1) to the second harmonic 2*f0.

Hi Mohan, the way to estimate threshold is usually follows the procedure described in behavioral (Psycho-physical) experiments.  You read about method of constant stimuli, that will provide you the possible answer. 

Vijay

Thanks for the suggestion. I checked and i am afraid there is no mention of tungsten epoxy backing material in that book. May i add this material is normally used as backing material in ultrasonic transducers or hydrophones.

It seems that 25kHz comes from the machine with it`s first harmonic at 50kHz. Just change the machine`s operating point [(rotational) speed, force, ...] to check if the peaks move in your FFT.

And it`s not clear what you mean with denoising: just filtering your sound waves or active noise control / reduction?

The question, that you asked, rises next problem: what does acoustic comfort mean? Do you mean noise free environment or absence of any acoustic signals? Or concert hall with good acoustic properties? If the last one, Ease could be good choice.

The second question: evaluate acoustic comfort for existing objects, or designed?

If you mean any existing building, measurements could be good starting point.

These are just a few quick thoughts on the question you asked.

Cavity noise is a commonly found in air crafts and air conditioning ducts.  This tend to be characterised by three main phenomenon: (Separation from the leading edge generates a shear flow that ends up creating (1) and (2))

1) A feedback loop inside the cavity due to recirculation.

2) An edge tone like phenomenon at the trailing edge.

3) Feedback pressure waves.

Such tones are now called the Rossiter modes,

I have solved this problem, because I have simply mistake, when I definite the PMLs, the center coordinate is not in according with the geometry . Thank you for your attention!

Hey Carlos,

just adding to the previous two answers: You can use the SPTOOL if you have the signal processing toolbox. The tool lets you view your signal and design filters and try them..

If you have a signal that has a temporal alignment, then use FILTFILT instead of FILTER to avoid phase shifts (e.g. use zero-phase filtering).

Greetings, David

The main ship parameters affecting wave propagation are the length, beam, draft, displacement and shape of the hull.

There are some secondary parameters too, like the smoothness of the surface: e.g. if there are sharp corners or other slope discontinuities, that can cause waves to be produced.

The roughness of the surface can increase turbulence close to the ship and might act to damp out very high frequency waves, especially immediately behind the hull.

I think you would get some excellent insights into the problem by looking at how wave resistance of model hulls is estimated from towing tank measurements. There are a number of methods (e.g. transverse cuts, longitudinal cuts, and the XY-method of Lawrence Ward) that you could look at.

About 20 years ago, I released a (free) version of the program "Michlet" which was a thin-ship, linear code that allowed the user to input a wave field behind the ship. The program used (primarily) memetic algorithms to search for the hull that produced that wave field.

Although there were severe restrictions in that toy problem, it was able to get reasonable estimates of the location and the length, beam, draft and displacement of the hull, as well as some indication of the shape of the hull.

Of course, the problem is ill-posed, and can be made even worse by adding real-world effects such as, among many others, ambient waves and viscous damping effects, but it can give some insights into the way various parameters act, and interact, in producing the far-field wave pattern.

You might use a finite element approach

Best regards,

                         Nils

What is your band requirement ?

What is the medium (air, water ...) ?

You can take a look to Bruel & Kjaer products, they are expensive but they are viewed by many as the best transducers (at least in my field, underwater acoustic).

For the acquisition device, poor chance a sound card (even a high-grade one) can do the job because they won't pass DC to 20Hz, 50Hz or 100Hz. The lowest I've found are from the RME brand, they are  advertising 5Hz at -3dB , and i've measured mine wich cuts at 1 or 2Hz.

That said, another culprit of doing acquisition with a sound board is if you need to synchronize the start of the recording session (no trigger input).

I'd suggest to experiment with cross-correlation techniques and try with different signals. Try with either constant frequency and modulated signals of different duration, e.g. short and long sweeps. Of course you have to manage the cross-correlation length accordingly.

Gianni

It dosent change the figure, it only had a effect on the far field microphone. I asked the Ansys support and they told me I should use an "Impendance Boundery" on the surface, with the impendance of the air. But i didn't tried it, because it will not effect the far field, only the acoustic body (the figure) you see. 

Conrad

Hi John,

With the hydrophone, you can measure the pressure in a single point, so you are measuring the pressure change in one point of the pressure field imposed by the sonotrode.

If you are working with just one bubble (state of single bubble cavitation) you would probably “see” the change in the shape of your hydrophone’s signal (that introduce the presence of the cavitation  bubble and if you use a high pass filter, applied to your hydrophone signal, you would probably see even easily the presence of a cavitation bubble), however if you are in a multi-bubble regime, you most probably see all the interaction of bubbles + the impose acoustic filed, and that’s is more difficult to work with, nevertheless you could build another high pas filter in order to see the cavitation in your chamber.

Another home-made way to measure the number of node in your chamber is by the use of paper metal strip, where you could see the damage in the strip that the cavitation bubbles made (when they collapse) 

Here is a good bibliography of the differences between the two regimens

Ole-Herman,

You still haven't told us why you want to use a hydrophone in air.

I agree with everyone on the impedance mismatch. Hydrophones just weren't designed to work in air.

If you are looking for a weatherproof or immersible microphone, there are a number available. Just search the web on "waterproof microphone". Some are fully immersible, although they probably don't work very well as hydrophones, just as hydrophones don't work very well in air.

A weatherproof mic in air will give you a much better signal-to-noise ratio and a much more predictable amplitude-frequency response and directional characteristic than a hydrophone in air.

If you want a transducer that will work in both media, I suggest you use both a hydrophone and a waterproof mic and mix them after the preamps. If you set the gains correctly, the impedance difference between the two media should effectively switch between the two transducers automatically.

As suggested by Prasad Kantipudi, start reading the literature, which is extremely rich and with plenty of tutorials. This topic has been investigated for decades and there are thousands of works. Today state of the art are i-vectors.

Note that you can find several free tools which implement these state-of-the art solutions.

During the first half of the 20th century, we measured RT by setting a stimulus such as octave-band-filtered noise to a certain level, then cutting off the stimulus and clicking a stopwatch at the same time. We then observed the decay on a sound level meter using the FAST setting and a D'Arsonval movement. When the decay reached 60 dB below the original steady-state sound, we clicked the stopwatch again. The stopwatch reading was the RT. Sabine used human-blown organ pipes as his sound source. The precision of these methods is limited by the human link in starting and stopping the stopwatch. The D'Arsonval meter smoothed the quasi-random amplitudes of the individual reflections making up the reverberant field, thus acting as sort of a mechanical Shroeder averager. This process can be improved by using a graphic level recorder or chart recorder to trace the sound level during the decay, but then, as Manuel points out, you have to do some sort of averaging. The pen speed setting on the recorder can do the averaging for you. I have measurements of a church sanctuary made using this method that gave RT results quite close to the results obtained using modern computer impulse-response-Schroeder methods.

Perhaps if I knew your application, and the reason for wanting to avoid Schroeder integration, I could be of more help.

(2*5m)/300m/s = 33msec for a soundwave travelling 5m towards and backwards to the transducer (puls-echo mode)

A complete measurement over 8 transducer for max. distance of 5m needs 267msec.

[So the puls repetition rate for each transducer will be 3,75]

Your robot is able to detect changes in distances of obstacles every 8/30s. Within this time at the given acceleration of a=0,5m/s^2 a maximum velocity of 0,133m/s (0,48km/h) will be reached.

But, if your robot is able to travel for say 10s at constant acceleration without collision the speed at the end will be 10s*0,5m/s^2 = 5m/s = 18km/h. Now at each ring cycle time of 267msec a distance of 1,33m will be covered.

At least you have to give a better definition of your maximum speed.

If it would be okay to avoid collisions in the range of one detection for 5m range then v = 5m / 0,267s = 18,75m/s. That means a speed of 67,5 km/h. Perhaps too fast for collision avoidance.

By the way: bandwidth and risetime is only interesting if you want to calculate time-of-flight. In this example the speed of the robot seems to be in the focus of interest.

Hi

Pzt transducers has one or more resonance frequencies, if you are in their resonance frequencies the performance of the transducer is maximal (at least in the frequency domain), so their intensity is also improved, however there are others ways to improve your transducer performance, such as “tuning the impedance of the transducer-medium”, and adjusting the frequency slightly in your experiment in order to have always the resonance frequency (yes, the resonance frequency has some little changes in course of time), if you want to know the resonance frequency of your transducer you should use an impedance analyzer

Thanks. The answers were helpful

Yichun,

You will need a signal acquisition deconvolution algorithm with paramerters determined by the signal acquisition system you are using.

Fundamentally, you can think of the process of signal blurring as the convolution of the  impulse response of the acquisition system with a "perfect" image. Then you deconvolve to remove the bluring.

You also have to have a spatial sampling rate that is adequate to model the Fourier Transform spectral components of the image. If you do not meet the Nyquist sampleing criterion, you data will have "spectral folding" and this contamination of the data cannot be removed.

So two things are required:

1. adequate spatial sampling

2. convolution model for the signal acquisition system.

Good Luck

@Dante, Thanks for the information and I will read the paper.

@Valerio, Thanks for your input. I am aware of the properties of different windows, but I still want to know the commonly used one in guided wave transmission coefficient calculation, as many papers just did not mention it. 

Dear Carlos Ariel Ferrer-Riesgo,

Thank you for the ideas.

I tried resampling the sound at 8 kHz and calculating  the integrals in the spectral domain. Though I received lower values than before, they are still orders of magnitude higher than those reported in the Fukazawa paper (they actually sampled at 20 kHz).

I still don't know what's going on, and, at this point, I don't think I'll use that measure in my study.

Sure, good luck.

Limited to stereo microphones you will have limited localization.  The source for the TDA algorithm is implemented in Java in the following code.

Dear Khalaf,

The music function aims at minimising the following cost function:

P_music(\theta) = ||u_1^H x a(\theta)||^2  +  ||u_2^H x a(\theta)||^2 + ... +  ||u_{N-q}^H x a(\theta)||^2 where u_1 u_2 till u_{N-q} are the eigenvectors corresponding to the smallest N-q eigenvalues L1 < L2 < .. L_{N-q} of data covariance matrix of size N and q being the number of sources.

EV is a weighted version of MUSIC weighting each term above of the form ||u_j^H x a(\theta)||^2 as follows  (1/Lj) ||u_j^H x a(\theta)||^2.

So minimising the above term is more reliable when weights are assigned, i.e. the smallest eigenvalues count more than larger ones, i.e. small eigenvalues are more reliable than large ones.

The first delay that you have is the buffer array size, that one is for example in you case  ( 1 / 44100 ) × buffer size. Then if you are polling are you sure you are polling fast enough for capturing at 44100 and second does the processing that you are doing is fast enough?  You could instead use interruptions to assure a 44100 sampling while your processing limit depends on your buffer size. Also measure your timings you could do this with the TI tools or with an oscilloscope, check your processing time your polling times. This could give you a good idea of what is happening.  

In addition, please check the normal direction defined in your model. Please make sure it is consistent with your BIE formulation.

I googled the probe you use and It seems to be ok for the job you are working on.  Did you use a good transmitting media, I mean ultrasound jel - plenty of it...  You may try to put ex vivo pig hearts in a plastic container filled with ultrasound gel, I have tried plain water for some other experiments, ultrasound gel may even work better.  

You can check my experimental model ( A model for basic ultrasound set-up and training for 3D/4D ultrasound scanning

K E Karaşahin · M Ercan · I Alanbay · U Keskin · M Dede · M C Yenen · I Başer

Ultrasound in Obstetrics and Gynecology 03/2011; 37(3):371-2. )

I believe it will work for you.  Submerge the pig heart  about 4- 5 cm away from the tip of the probe, make sure there is plenty of gel between the heart and the probe.  Try to avoid air bubbles. 

You can also try to fill the plastic cup with jello solution, and observe after it solidifies.  

4 D imaging is real time so you have to be patient, do not move the probe around quickly.  3 D is easier but still requires a stationary probe.  Get a clear 2 D image first.  Then use 3/4 D .

Use small volume or interest for better results (you get higher frequency and better resolution).

Good luck.  

APC has a nice online calculator for calculating the frequency based on the material type and dimensions.

The short answer is no. Formant frequencies are variable and context-dependent even within speaker. One well-known reference is Hildebrand et al. JASA 1995. The long answer is that with a little more effort, it is possible to do pretty well provided that you identify beforehand the vowel core. See for example http://www.ling.upenn.edu/courses/cogs501/Hillenbrand.html for a modern machine-learning approach to the Hildebrand data.

Hi,

Yes I have been looking at RNN-RBMs, even overviewed Nicolas Boulanger-Lewandowski's thesis but it's all about generating *scores* which is very different from generating actual sound.

I will look into LSTM and GRU. Thanks.

Hi Arpit,

Assuming that you are using your 2 microphones as a ULA (Uniform Linear Array), the separation of 6cm/0.06m will correspond to a wavelength of 0.12m.  Assuming the speed of sound is 343m/s, the wavelength of 0.12m will correspond to a frequency of 2.858 kHz.  I am thinking that if your are sampling a audio signal greater than 2.858 kHz, you may want to place the microphones nearer to each other (less than 6 cm, maybe 5 cm for up to 3 kHz) so as to prevent signal returns due to grating lobes from interfering with your intended audio signals. 

Also, when using MVDR, an array of 2 elements with only provide two degrees of freedom (DOF) for the constraints: (1) to prevent distortion of the signal in the desired direction of arrival and the (2) to suppress interference from other directions.  Thus, by adding a couple more of array elements (microphones), you will be able to produce better interference suppression from undesirable signals from other directions.

See the attached file.

Dr Ruiter,

Seeing your affiliation reminds me that in the early 1980's I used empirical formulas published by V.M.A. Peutz for predicting the performance of sound reinforcement systems in rooms. After comparing those equations to standard reverberation equations, I came to realize that Peutz was treating room reverberation as another noise term.  Modern acoustic analysis instruments can compute a variety of speech intelligibility measures, but %ALcons continues to be widely used by practicing sound contractors. 

There are many versions of what you're talking about. 

Example, a root EV approach, close to ROOT MUSIC.

Could you be more specific ? 

For score-informed source separation like the reference indicated, it is a semi-blind source separation.

Hi, i guess you need a proper Data Acquisition Device (DAQ).

During my bachelor works, we did use NI-DAQ USB-6001 with 13 digital I/O and 8 analog inputs. With a proper configuration of Matlab Data Acquisition toolbox, you can reach almost real-time processing (which in our case was adequate enough).

Cheers!

Hi,

PhysioNet, probably one of the best source for physiological data actually provides a database which includes OAE data:

I definitely recommend this site to everyone who wants to work with physiological data. I hope it helps you.

Best,

Johannes