Hi

Your question is very wide. My answer below in not intended to discourage you - merely, to save you time in finding/searching out what it is that you actually need.

To start with, you need to distinguish between high and low frequency, i.e. between object size that is larger or smaller than the wavelength, respectively.

High frequency models, e.g. Statistical Energy Analysis models, are small and fast but often not very accurate. Low frequency models may be more accurate - up to a point - where Amplitude/Phase no longer is reliable, which is when you must turn to high frequency mode of reasoning whether you like it or not.

You state that a 10% error margin vill suffice, i.e. an error of about 1 dB or less will suffice. It is well known from high frequency methodology that the amplitude error is 5-6 dB per standard deviation, i.e. +/- 10-12 dB with 90% probability.

The high frequency limit is referred to as the Schröder frequency. This limit is well known in acoustics and was explained in 1954 by M.A. Schröder as the condition (for rooms) where you have three natural frequencies (modes) overlapping within the half power bandwidth of a mode. This is referred to as a condition with high modal overlap.

As you may imagine, phase rapidly shifts 180 degrees when frequency is changed. Likewise, when the system changes and thereby shifts the resonance frequency you get a large phase shift. The net result is random phase and summation of three modal contributions (view these as three vectors of equal magnitude and random orientation in the Nyquist space) add upp to a magnitude variation of 5-6 dB/standard deviation.

The Schröder frequency is hit for most systems at some frequency. Differently put, it is not a question of if but of when.

http://www.code-aster.org/forum2/attachment.php?item=7045

http://www.akutek.info/Papers/MS_Schroeder_Revisited.pdf

https://en.wikipedia.org/wiki/Room_acoustics

You may protest that submerged may imply unbounded domains. OK, but you will then be exposed to things like temperature gradients and/or flow which upset the assumption on uniform (or at least known) fluid properties which most BE/FE models rely on. This limit will not be the formal Schröder frequency but it will involve the same physics reasoning.

So, to sum up - if your question refers to a small object submerged into still standing water, you can likely analyse of up many kHz. If your object is a large structure, say an oil platform, you will hit the Schröder frequency at a few Hz.

Next, calculation method. As a rule of thumb, BE generates fully populated, hard to solve, matrices while FE generates larger, sparser, matrices that are faster to solve. There exists infinite finite elements that work well. 

The main attraction to BE used to be simpler meshing. I believe this no longer is the case. That said the infinite finite elements do not extend beyond your analysis boundary. The FE analysis model size grows with the cube of the relative distance away from the object is is required to cover.

I define relative distance as distance/wavelength, i.e. close  as small and far away as large relative distance values.

So, if you want to analyse sound close to a submerged structure you likely are better off with FE but BE would be your method of choise when response should be calculated far away.

Submerged structures should incorporate the influence of fluid loading. This can be done with weak or strong formulations, i.e. with formulations where the structure ignores its effect on the fluid (or vice versa) and where the two communicate, respectively.

The above reasoning can be extended to cover also a variety of other things you may need to consider for your model.

The message I am trying to put across is that the heaviest methods or the strictest modelling assumptions are not always called for as one always must weigh it against the Schröder frequency as well as the accuracy of the modelling assumptions and the purpose for which your model is made.

Therefore, in my opinion - there is no general 'best method' as both method and modelling assumptions must adapt to what can be achieved and what can be motivated in your modelling/computation effort.

If you are interested, you find some ramblings on the topic of modelling and model QA on my home page, e.g. here

http://qringtech.com/learnmore/why-simulate-measure-correlate-automate/

http://qringtech.com/wp-content/uploads/2014/01/AnalysisTechnology.pdf

Here is discussed a mix of low/high frequency methodology

http://qringtech.com/2010/09/15/vibroacoustic-fe-analysis-piping-vibration-transmission/

Hope this helps

Claes