Why the numbers look small

Full scale and dBFS

A 16-bit sample runs from −32768 to 32767. dBFS measures a signal against that full scale: 0 dBFS is the biggest a sample can be, and every 20 dB down is ten times smaller.

The datasheet gives one number that ties sound to samples: −26 dBFS at 94 dB SPL, typical, within ±1 dB. 94 dB SPL is the test level, about a loud voice a few centimetres from the port. So that loud voice reads about a twentieth of full scale.

Speech is about a thousandth

Every dB of sound is a dB of signal. Normal speech at arm's length is about 60 dB SPL, 34 dB quieter than the test level, so it lands near −60 dBFS: about a thousandth of full scale, a peak of a few tens out of 32767. That is worked out from the datasheet, not measured, and it is why a working microphone prints numbers that look broken.

At the bottom, the microphone's signal-to-noise ratio is 58 dB, so anything below about 36 dB SPL is as quiet as its own hiss. At the top, it overloads at 120 dB SPL.

Two corrections every sketch makes

• Subtract the offset. Decimated samples sit on a small constant offset, so silence is not exactly 0. The sketches take the average of each block and subtract it; left in, the offset reads as sound that is not there.
• Multiply by a gain. Speech a few tens high is barely audible played straight to a DAC. The live sketch multiplies by 4, then constrain() keeps the result inside 16 bits, because anything past 32767 would wrap round to a large negative number: a loud click.

The level meter in How loud is the room does the first and prints the result without gain, so you see the raw size.