Q1) Is it possible to use generated signal that passes before the power amplifier to obtain impulse response??!
Maybe. The main problem here is that both your speaker voltage and acquired microphone signal are bandpass signals: you only excite from 20Hz to 20kHz, there is AC coupling and anti aliasing filters. That means below 20Hz and above 20 kHz you have very poor signal to noise ratio and if you try to unwind the excitation you can get a huge amount of noise amplification. This typically requires some carefully manual management and in general the calculated impulse response will have pre-ringing (i.e. isn't as causal as the real physical system).
(It would be nice if this was possible, but it is expected that the changes in the amplitude and phase of the signal transmitted to the speaker by the amp. should be reflected).
This is odd to me. You are using a very high quality (and very expensive) measurement amplifier. This should be flat as a pan cake and the smallest source of error in your system. I suspect that the phase difference you see, is really just a small frequency independent delay of class D modulator in the amp. Interestingly enough B&K doesn't spec the phase response or latency of the amp, which given the price of that thing is almost criminal negligence :-)
Q2) If it is not feasible, Is there a good way to get IR using the measured input voltage (From power amp.) and response(pressure)?
I personally prefer to use an excitation that has a well behaved inverse across the entire spectrum (from DC to Nyquist) and keep the effects of amp, A/D and D/A as part of the system. They are hard to unwind and if you use half way decent gear (as you) the impact on the measurement is minimal
Q3) In order to obtain the IR of the room, the Input is set to Voltage, but in my opinion,the Input should be the volume acceleration of the sound source or a physical quantity corresponding to it. Is there any reason why it is OKAY to calculate IR like this?
The things that you are measuring are the input voltage to the speaker and the output of the microphone which are indeed a voltages. What acoustic quantity that corresponds to, depends on the microphone & speaker. For the microphone it's either pressure or particle/volume velocity (not acceleration). As long as you are reasonably far away from a reflecting boundary these are proportional and related through the free field impedance of air.
The speaker turns the input voltage into a very complicated 3-dimensional pressure field. Trying to define a single pressure reference signal is a tricky. For direct radiators you can use "anechoic at 1m" but for the type of speaker that you are using that would be hard to justify (in my opinion), so voltage seems the best choice. At least it's well defined and reproducible.
You can certainly calibrate the entire setup, if you want to.
Q4) Are there any other considerations other than those mentioned above to get an accurate IR?
Measuring room impulse response accurately is VERY difficult. Things that need to be carefully managed and assessed are
- Signal to noise ratio as a function of frequency. Most rooms are very noisy especially at low frequencies. There is HVAC, building vibrations, traffic noise, squeaky chairs, people talking, etc. All these needs to be monitored. Your measurement is only going to be "good" over a certain frequency range and you need to determine what the range is, so you can remove the other crap.
- Non-linear distortions of the loudspeakers. All electromagnetic loudspeakers are somewhat non-linear and the non linearity increases with level. Finding the best excitation level is hard. Too low gives you poor SNR and too high gives a lot of THD.
- Placement and time variance. You need to pick good spots and make sure nothing moves during the measurement.
The devious part is that none of those errors can be detected by "visual inspection". The measurement looks like a perfectly good impulse response even if it's full of actual measurement errors. So ideally you wrap a set of diagnostics around the whole process that quantifies each individual error mechanism and makes a "good/bad" call based on the requirements of your specific application. What errors can you tolerate and to what amount?
In the end you need to verify your whole setup: how can you do that? How do you know what you measured is actually "correct" or at least "good enough"?
Fortunately if all you want in the end is reverb time, this shouldn't be too hard. I strongly recommend using Schroeder integration, which is rather forgiving in terms of measurement errors.
Update on noise amplification
Your measured signal, $M(\omega)$ always contains some amount of noise., i.e.
$$M(\omega) = Y(\omega)+N(\omega) = H(\omega)X(\omega)+N(\omega)$$
where $H$ is the transfer function, $X$ the input signal and $N$ the noise spectrum (acoustic noise, line hum, etc) . You can estimate the transfer function by dividing the FFTs:
$$\hat{H}(\omega) = \frac{Y(\omega)+N(\omega)}{X(\omega)} = \frac{H(\omega)X(\omega)+N(\omega)}{X(\omega)} = H(\omega) + \frac{N(\omega)}{X(\omega)} $$
So if you have no energy at $10Hz$ in your excitation signal but a fair amount of noise at $10Hz$ you'll end up with a lot of $10Hz$ in your measured impulse response. Dividing something by a small number makes it very big.