What is the best way to implement a simple energy-efficient compact voice recorder (device)?

Question

Do it yourself

It is necessary to design a relatively cheap device capable of continuously recording dozens of hours of sound (voice) from a single charge. At the same time, it should be compact and light enough to be worn as a clip on clothes.
It is important that the voices of the owner and people in the immediate environment of 2-3 meters are clearly distinguishable.
Specifically, I would like to solve the following issues:

1. Should I use a codec (software or hardware) for compression, or is it more efficient to write PCM stupidly, since flash drives are now cheap and bulky? The format is not important, post-processing will solve everything.
   
2. Does it make sense to bother skipping silence in terms of energy efficiency? I don't want to lose the beginning of the phrase, do I have to use some kind of circular buffer?
   
3. Is it possible to add the possibility of episodic access to fragments of the recorded sound via BlueTooth through your application from a smartphone to the conditions voiced above?
   
4. Is it better to use a controller ADC for digitization or is it better to take a microphone with SPI?
   
5. Which microphone is better to use, is it difficult to make it strapping for the given requirements?
   
6. Which controller to use depending on the previous options? TI, Atmel, STM?..
   
7. Solder memory or use TF Card?
   

Interested in general considerations, estimates, options ...

Answer 1

1. use the simple ADPCM codec, its quality is quite enough for voice recording. and the flow compression will turn out to be quite good.
reducing the amount of writing to a flash drive will increase the operating time (writing to a flash also consumes energy well)
2. has. the algorithm is simple. the result is excellent. then you have flash drives that are dimensionless to the point of not giving a damn, then you are afraid to take more ram. writing hours of silence is at least stupid and wasteful.
3. everything is real. what you write will be, you just have to sit down to write and finish it to the mind.
4. The ADC of the controller is enough for the voice, an ADC of the same not high quality is shoved into a cheap SPI microphone. in general, test the record and the solution will be clear to you.
5. any suitable for your conditions: directional / omnidirectional, frequency response sensitivity ...
they are all almost now capacitive with a built-in field worker. a large external strapping is not required.
6. any suitable. write down what you want and get the preliminary parameters of the iron.
7. use a tf card - they are now like dirt, and the cost is appropriate.