This is the solution I think would work, there might be other more accurate and efficient ways to do this.
You might do this steps for every recording:
- Find the silent sections (using the algorithms stated below)
- Measure Standard Deviation of those silent portions of your recordings as an indication of noise power.
Select Standard Deviation of silence sections of one the recordings that you are sure has acceptable quality, as a your threshold value.
Discard any other recording that its estimated noise power is higher that specified threshold (which you estimate in step 3).
For detection of silent portions of the voice, I suggest simply segment your signal into some non-overlapping windows, estimate signal power
any signal segment which has a signal power below a per-defined threshold, might be considered as silent section.
However this method is not very accurate. If you have enough time, try an algorithm called Voice Activity Detection (VAD), which is more widely used (the ITU-T G.729 standard uses VAD to reduce the transmission rate during silence periods of speech). Please, take a look at http://practicalcryptography.com/miscellaneous/machine-learning/voice-activity-detection-vad-tutorial/