diff --git a/src/rtl_sdr_fm.cpp b/src/rtl_sdr_fm.cpp
index 2c52443..1906b1e 100644
--- a/src/rtl_sdr_fm.cpp
+++ b/src/rtl_sdr_fm.cpp
@@ -943,28 +943,27 @@ static inline void audio_lpf_filter(struct demod_state *fm)
     fm->audio_lpf_state = y;
 }
 
-long int rms(int16_t *samples, int len, int step)
-/* largely lifted from rtl_power */
+long int mean_power(int16_t *samples, int len, int step)
+/* DC-corrected mean power (sqrt-free). Returns squared RMS units. */
 {
 	int i;
-	long int rms;
-	long p, t, s;
+	int64_t p = 0;
+	int64_t t = 0;
+	int64_t s;
 	double dc, err;
 
-	p = t = 0L;
-	for (i=0; i<len; i+=step) {
-		s = (long)samples[i];
+	for (i = 0; i < len; i += step) {
+		s = (int64_t)samples[i];
 		t += s;
 		p += s * s;
 	}
 	/* correct for dc offset in squares */
-	dc = (double)(t*step) / (double)len;
-	err = t * 2 * dc - dc * dc * len;
+	dc = (double)(t * step) / (double)len;
+	err = (double)t * 2.0 * dc - dc * dc * (double)len;
 
-	rms = (long int)sqrt((p-err) / len);
-	//going with a value that's easy to figure out when its done the thing
-	if (rms < 0){ rms = 999; }
-	return rms;
+	double power = ((double)p - err) / (double)len;
+	if (power < 0.0) power = 0.0;
+	return (long int)power;
 }
 
 void full_demod(struct demod_state *d)
@@ -1939,8 +1938,9 @@ long int rtl_return_rms()
 	//debug -- on main machine, lp_len is around 6420, so this probably contributes to very high CPU usage
 	// fprintf (stderr, "LP_LEN: %d \n", demod.lp_len);
 	//I've found that just using a sample size of 160 will give us a good approximation without killing the CPU
-	// sr = rms(demod.lowpassed, demod.lp_len, 1);
-	sr = rms(demod.lowpassed, 160, 1); //I wonder what a reasonable value would be for #2 (input len) there
+	// Return mean power (squared RMS) for soft squelch decisions (sqrt-free)
+	// sr = mean_power(demod.lowpassed, demod.lp_len, 1);
+	sr = mean_power(demod.lowpassed, 160, 1);
 	// #endif
 	return (sr);
 }