There is a bug in the "silence" effect that appears to
have been introduced in 2009 with commit bbb403.
In function aboveThreshold() the value being scaled
is effectively the result of a sqrt but when it
gets scaled the value is not rounded but coarsely truncated instead.
Truncation is wrong. Rounding is right.
Prior to this commit the value was effectively
rounded to the nearest whole value (according to
the precision). Without the following patch the "silence"
results are considerably different to earlier versions.
For example:
value old masked current
----- ---------- -------
0x009bd6ae 0x0000009c 0x009b0000
0x00a170b8 0x000000a1 0x00a10000
The 0x009bd6ae used to be correctly rounded to 0x9c whereas
the current version effectively truncates the value to 0x9b.
That is a serious flaw that should have been picked up
during testing. But better discovered 11 years later by a
random member of the public than never.
The patch:
======================================================================
--- sox-downstream-sox-14.4.2.0.modified/src/silence.c.jw
+++ sox-downstream-sox-14.4.2.0.modified/src/silence.c
@@ -277,9 +277,15 @@
{
/* When scaling low bit data, noise values got scaled way up */
/* Only consider the original bits when looking for silence */
- sox_sample_t masked_value = value & (-1 << (32 - effp->in_signal.precision));
+ double scaled_value;
+ sox_sample_t rounded_value = value;
- double scaled_value = (double)masked_value / SOX_SAMPLE_MAX;
+ /* before we mask we should round the value */
+ if (effp->in_signal.precision < 32)
+ rounded_value += (1 << (32 - effp->in_signal.precision - 1));
+ rounded_value &= (-1 << (32 - effp->in_signal.precision));
+
+ scaled_value = (double)rounded_value / SOX_SAMPLE_MAX;
if (unit == '%')
scaled_value *= 100;
======================================================================
:JW
_______________________________________________
SoX-devel mailing list
SoX-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/sox-devel
