Reducing the bit width reduces the sound quality, which degrades the performance of the recognition, as does reducing the sample rate, which means limiting the bandwidth. 4KHz bandwidth, 8 KHz sample rate covers all of speech, but not singing.
I don't know how much it'll be degraded by using 15 or 14 bits... there's be a lot of math in fixing the serial data bytes into 14 or 15 bits. 8 would probably be a huge problem.
The resolution you get with a particular bit width is related to the dynamic range used. For example, if your ADC is configured so that the max and min 16-bit codes occur with 1vpp, but the microphone is amplified so that speech only produces maximum vpp of 250mV, then that's the same resolution of 14 bit sound that goes through the entire bit space. But, it has the headroom to accommodate louder speech if needed, and clipping from experiencing sound louder than the ADC has a code for is bad. Getting just the right amplification is actually a bit tricky. The problem is MUCH worse with open mikes, because the sound level varies a lot with distance.