Menu
▾
▴
iaxclient-cvs
|
From: <jpg...@us...> - 2007-05-22 12:17:02
|
Revision: 980
http://svn.sourceforge.net/iaxclient/?rev=980&view=rev
Author: jpgrayson
Date: 2007-05-22 05:17:03 -0700 (Tue, 22 May 2007)
Log Message:
-----------
Fix problem with the visual studio compiler not appreciating the magnitude of the floating point constant -9e99f.
Modified Paths:
--------------
trunk/lib/audio_encode.c
Modified: trunk/lib/audio_encode.c
===================================================================
--- trunk/lib/audio_encode.c 2007-05-21 21:37:10 UTC (rev 979)
+++ trunk/lib/audio_encode.c 2007-05-22 12:17:03 UTC (rev 980)
@@ -25,7 +25,7 @@
#include "codec_ilbc.h"
#endif
-float iaxc_silence_threshold = -9e99f;
+float iaxc_silence_threshold = -99.0f;
static float input_level = 0.0f;
static float output_level = 0.0f;
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
|
From: <sb...@us...> - 2007-09-10 17:52:47
|
Revision: 1117
http://iaxclient.svn.sourceforge.net/iaxclient/?rev=1117&view=rev
Author: sbalea
Date: 2007-09-10 10:52:49 -0700 (Mon, 10 Sep 2007)
Log Message:
-----------
Fix db calculation for sound levels to return -99.9 when there's no audio
Modified Paths:
--------------
trunk/lib/audio_encode.c
Modified: trunk/lib/audio_encode.c
===================================================================
--- trunk/lib/audio_encode.c 2007-09-05 23:31:26 UTC (rev 1116)
+++ trunk/lib/audio_encode.c 2007-09-10 17:52:49 UTC (rev 1117)
@@ -58,7 +58,7 @@
/* avoid calling log10() on zero which yields inf or
* negative numbers which yield nan */
if ( vol <= 0.0 )
- return 0.0;
+ return -99.9;
else
return log10f(vol) * 20;
}
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
|
From: <jpg...@us...> - 2007-10-03 17:15:49
|
Revision: 1165
http://iaxclient.svn.sourceforge.net/iaxclient/?rev=1165&view=rev
Author: jpgrayson
Date: 2007-10-03 10:15:52 -0700 (Wed, 03 Oct 2007)
Log Message:
-----------
Be explicit about float use.
Modified Paths:
--------------
trunk/lib/audio_encode.c
Modified: trunk/lib/audio_encode.c
===================================================================
--- trunk/lib/audio_encode.c 2007-10-02 19:37:29 UTC (rev 1164)
+++ trunk/lib/audio_encode.c 2007-10-03 17:15:52 UTC (rev 1165)
@@ -57,8 +57,8 @@
{
/* avoid calling log10() on zero which yields inf or
* negative numbers which yield nan */
- if ( vol <= 0.0 )
- return -99.9;
+ if ( vol <= 0.0f )
+ return -99.9f;
else
return log10f(vol) * 20;
}
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
|
From: <sb...@us...> - 2007-11-07 15:43:18
|
Revision: 1271
http://iaxclient.svn.sourceforge.net/iaxclient/?rev=1271&view=rev
Author: sbalea
Date: 2007-11-07 07:43:16 -0800 (Wed, 07 Nov 2007)
Log Message:
-----------
Speex changed its interface to VAD probabilities. In early 1.1 dev versions
VAD probabilities were floats. In 1.2beta1, the probabilities are 32 bit ints.
Iaxclient still passed pointers to floats as arguments, and this forced speex
to use its built in defaults. In effect our VAD probability tweaks did nothing.
This change attempts to fix this by converting VAD probability arguments to
32 bit integers ranging from 0 to 100.
Modified Paths:
--------------
trunk/lib/audio_encode.c
Modified: trunk/lib/audio_encode.c
===================================================================
--- trunk/lib/audio_encode.c 2007-11-06 18:29:21 UTC (rev 1270)
+++ trunk/lib/audio_encode.c 2007-11-07 15:43:16 UTC (rev 1271)
@@ -94,7 +94,6 @@
static void set_speex_filters()
{
int i;
- float f;
if ( !st )
return;
@@ -107,10 +106,10 @@
speex_preprocess_ctl(st, SPEEX_PREPROCESS_SET_DENOISE, &i);
/* make vad more sensitive */
- f = 0.30f;
- speex_preprocess_ctl(st, SPEEX_PREPROCESS_SET_PROB_START, &f);
- f = 0.07f;
- speex_preprocess_ctl(st, SPEEX_PREPROCESS_SET_PROB_CONTINUE, &f);
+ i = 30;
+ speex_preprocess_ctl(st, SPEEX_PREPROCESS_SET_PROB_START, &i);
+ i = 7;
+ speex_preprocess_ctl(st, SPEEX_PREPROCESS_SET_PROB_CONTINUE, &i);
}
static void calculate_level(short *audio, int len, float *level)
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
|
From: <sb...@us...> - 2008-02-06 00:08:55
|
Revision: 1343
http://iaxclient.svn.sourceforge.net/iaxclient/?rev=1343&view=rev
Author: sbalea
Date: 2008-02-05 16:09:00 -0800 (Tue, 05 Feb 2008)
Log Message:
-----------
Do not use const for variables that actually change.
Modified Paths:
--------------
trunk/lib/audio_encode.c
Modified: trunk/lib/audio_encode.c
===================================================================
--- trunk/lib/audio_encode.c 2008-02-06 00:00:52 UTC (rev 1342)
+++ trunk/lib/audio_encode.c 2008-02-06 00:09:00 UTC (rev 1343)
@@ -168,7 +168,7 @@
if ( (i & 0x3f) == 0 )
{
- const float loudness;
+ float loudness;
speex_preprocess_ctl(st, SPEEX_PREPROCESS_GET_AGC_LOUDNESS, &loudness);
if ( loudness > 8000.0f || loudness < 4000.0f )
{
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
|
From: <jpg...@us...> - 2008-04-15 22:33:30
|
Revision: 1411
http://iaxclient.svn.sourceforge.net/iaxclient/?rev=1411&view=rev
Author: jpgrayson
Date: 2008-04-15 15:33:37 -0700 (Tue, 15 Apr 2008)
Log Message:
-----------
Fix prototypes of some internal functions to reduce casting.
Modified Paths:
--------------
trunk/lib/audio_encode.c
Modified: trunk/lib/audio_encode.c
===================================================================
--- trunk/lib/audio_encode.c 2008-04-15 22:23:20 UTC (rev 1410)
+++ trunk/lib/audio_encode.c 2008-04-15 22:33:37 UTC (rev 1411)
@@ -118,7 +118,7 @@
speex_preprocess_ctl(st, SPEEX_PREPROCESS_SET_PROB_CONTINUE, &i);
}
-static void calculate_level(short *audio, int len, float *level)
+static void calculate_level(const short *audio, int len, float *level)
{
int big_sample = 0;
int i;
@@ -133,7 +133,7 @@
*level += ((float)big_sample / 32767.0f - *level) / 5.0f;
}
-static int input_postprocess(void *audio, int len, int rate)
+static int input_postprocess(short * audio, int len, int rate)
{
static float lowest_volume = 1.0f;
float volume;
@@ -149,7 +149,7 @@
set_speex_filters();
}
- calculate_level((short *)audio, len, &input_level);
+ calculate_level(audio, len, &input_level);
/* go through the motions only if we need at least one of the preprocessor filters */
if ( (iaxci_filters & (IAXC_FILTER_DENOISE | IAXC_FILTER_AGC | IAXC_FILTER_DEREVERB)) ||
@@ -213,9 +213,9 @@
return volume < iaxci_silence_threshold;
}
-static int output_postprocess(void *audio, int len)
+static int output_postprocess(const short * audio, int len)
{
- calculate_level((short *)audio, len, &output_level);
+ calculate_level(audio, len, &output_level);
do_level_callback();
@@ -342,8 +342,8 @@
/* decode encoded audio; return the number of bytes decoded
* negative indicates error */
-int audio_decode_audio(struct iaxc_call * call, void * out, void * data, int len,
- int format, int * samples)
+int audio_decode_audio(struct iaxc_call * call, void * out, void * data,
+ int len, int format, int * samples)
{
int insize = len;
int outsize = *samples;
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
|
From: <jpg...@us...> - 2008-07-01 15:35:17
|
Revision: 1436
http://iaxclient.svn.sourceforge.net/iaxclient/?rev=1436&view=rev
Author: jpgrayson
Date: 2008-07-01 08:35:07 -0700 (Tue, 01 Jul 2008)
Log Message:
-----------
Fix AAGC.
- Previously it was very common for the AAGC code to drive the input mixer
level (input volume) very low and potentially take a long time to recover.
- The new AAGC heuristic takes into account the amount of time needed for
the speex-provided "loudness" value to recover after a loud event.
- Also fixup input_postprocess() so that speex_preprocess_run() is called
with the audio_lock held.
Modified Paths:
--------------
trunk/lib/audio_encode.c
Modified: trunk/lib/audio_encode.c
===================================================================
--- trunk/lib/audio_encode.c 2008-07-01 15:21:20 UTC (rev 1435)
+++ trunk/lib/audio_encode.c 2008-07-01 15:35:07 UTC (rev 1436)
@@ -41,7 +41,7 @@
#include "codec_ilbc.h"
#endif
-float iaxci_silence_threshold = AUDIO_ENCODE_SILENCE_DB;
+static float iaxci_silence_threshold = AUDIO_ENCODE_SILENCE_DB;
static float input_level = 0.0f;
static float output_level = 0.0f;
@@ -51,7 +51,11 @@
static SpeexPreprocessState *st = NULL;
static int speex_state_size = 0;
static int speex_state_rate = 0;
-int iaxci_filters = IAXC_FILTER_AGC|IAXC_FILTER_DENOISE|IAXC_FILTER_AAGC|IAXC_FILTER_CN;
+static int iaxci_filters =
+ IAXC_FILTER_AGC |
+ IAXC_FILTER_DENOISE |
+ IAXC_FILTER_AAGC |
+ IAXC_FILTER_CN;
static MUTEX audio_lock;
@@ -72,16 +76,6 @@
static char outRingBuf[EC_RING_SIZE];
#endif
-/* AAGC threshold */
-#define AAGC_VERY_HOT 16
-#define AAGC_HOT 8
-#define AAGC_COLD 4
-
-/* AAGC increments */
-#define AAGC_RISE_SLOW 0.10f
-#define AAGC_DROP_SLOW 0.15f
-#define AAGC_DROP_FAST 0.20f
-
/* use to measure time since last audio was processed */
static struct timeval timeLastInput ;
static struct timeval timeLastOutput ;
@@ -178,11 +172,17 @@
static int input_postprocess(short * audio, int len, int rate)
{
- static float lowest_volume = 1.0f;
- float volume;
- int silent = 0;
+ static int aagc_frame_count = 0;
+ static int aagc_periods_to_skip = 0;
+ const int using_vad = iaxci_silence_threshold > 0.0f;
+ const int aagc_period = rate / len; /* 1 second */
+
+ int speaking = 1;
+ int loudness = 0;
+
MUTEXLOCK(&audio_lock);
+
if ( !st || speex_state_size != len || speex_state_rate != rate )
{
if (st)
@@ -199,75 +199,114 @@
i = ECHO_SUPPRESS_ACTIVE;
speex_preprocess_ctl(st, SPEEX_PREPROCESS_SET_ECHO_SUPPRESS_ACTIVE, &i);
}
-#endif
+#endif
speex_state_size = len;
speex_state_rate = rate;
set_speex_filters();
}
+
+ /* go through the motions only if we need at least one of the
+ * preprocessor filters */
+ if ( using_vad || (iaxci_filters &
+ (IAXC_FILTER_DENOISE |
+ IAXC_FILTER_AGC |
+ IAXC_FILTER_DEREVERB |
+ IAXC_FILTER_ECHO)) )
+ {
+ speaking = speex_preprocess_run(st, audio);
+ speex_preprocess_ctl(st, SPEEX_PREPROCESS_GET_AGC_LOUDNESS,
+ &loudness);
+ }
+
MUTEXUNLOCK(&audio_lock);
- calculate_level(audio, len, &input_level);
+ /* If we are using the VAD test and if speex indicates non-speaking,
+ * ignore the computed input level and indicate to the user that the
+ * input level was zero.
+ */
+ if ( using_vad && !speaking )
+ input_level = 0.0f;
+ else
+ calculate_level(audio, len, &input_level);
- /* go through the motions only if we need at least one of the preprocessor filters */
- if ( (iaxci_filters & (IAXC_FILTER_DENOISE | IAXC_FILTER_AGC | IAXC_FILTER_DEREVERB | IAXC_FILTER_ECHO)) ||
- iaxci_silence_threshold > 0.0f )
- silent = !speex_preprocess(st, (spx_int16_t *)audio, NULL);
-
- /* Analog AGC: Bring speex AGC gain out to mixer, with lots of hysteresis */
- /* use a higher continuation threshold for AAGC than for VAD itself */
- if ( !silent &&
- iaxci_silence_threshold != 0.0f &&
- (iaxci_filters & IAXC_FILTER_AGC) &&
- (iaxci_filters & IAXC_FILTER_AAGC)
- )
+ /* Analog Automatic Gain Control, AAGC. */
+ if ( speaking && iaxci_silence_threshold != 0.0f &&
+ (iaxci_filters & IAXC_FILTER_AGC) &&
+ (iaxci_filters & IAXC_FILTER_AAGC) &&
+ ++aagc_frame_count % aagc_period == 0 &&
+ !aagc_periods_to_skip-- )
{
- static int i = 0;
+ /* This heuristic uses the loudness value from the speex
+ * preprocessor to determine a new mixer level. The loudness
+ * ranges from 0 to up over 80. When mixer level, speex AGC,
+ * and the actual speaker's level are in equilibrium, the
+ * loudness tends to be from 4 to 16. When the loudness goes
+ * above this comfortable range, there is a risk of the input
+ * signal being clipped. AAGC's primary purpose is to avoid
+ * clipping.
+ *
+ * After a loud event (think cough), the loudness level will
+ * spike and then decay over time (assuming the speaker
+ * speaking at a relatively constant level). To avoid
+ * over-adjusting, we skip some number of aagc sampling periods
+ * before making any more adjustments. This gives the loudness
+ * value time to normalize after one-time spikes in the input
+ * level.
+ */
- i++;
+ /* The mixer level is a percentage ranging from 0.00 to 1.00 */
+ const float mixer_level = iaxc_input_level_get();
+ float new_mixer_level = mixer_level;
- if ( (i & 0x3f) == 0 )
+ if ( loudness > 40 )
{
- int loudness;
- speex_preprocess_ctl(st, SPEEX_PREPROCESS_GET_AGC_LOUDNESS, &loudness);
- if ( loudness > AAGC_HOT || loudness < AAGC_COLD )
- {
- const float level = iaxc_input_level_get();
+ new_mixer_level -= 0.20f;
+ aagc_periods_to_skip = 8;
+ }
+ else if ( loudness > 25 )
+ {
+ new_mixer_level -= 0.15f;
+ aagc_periods_to_skip = 4;
+ }
+ else if ( loudness > 15 )
+ {
+ new_mixer_level -= 0.10f;
+ aagc_periods_to_skip = 2;
+ }
+ else if ( loudness > 12 )
+ {
+ new_mixer_level -= 0.05f;
+ aagc_periods_to_skip = 4;
+ }
+ else if ( loudness < 2 )
+ {
+ new_mixer_level += 0.15f;
+ aagc_periods_to_skip = 4;
+ }
+ else if ( loudness < 4 )
+ {
+ new_mixer_level += 0.10f;
+ aagc_periods_to_skip = 4;
+ }
+ else
+ {
+ aagc_periods_to_skip = 0;
+ }
- if ( loudness > AAGC_VERY_HOT && level > 0.5f )
- {
- /* lower quickly if we're really too hot */
- iaxc_input_level_set(level - AAGC_DROP_FAST);
- }
- else if ( loudness > AAGC_HOT && level >= 0.15f )
- {
- /* lower less quickly if we're a bit too hot */
- iaxc_input_level_set(level - AAGC_DROP_SLOW);
- }
- else if ( loudness < AAGC_COLD && level <= 0.9f )
- {
- /* raise slowly if we're cold */
- iaxc_input_level_set(level + AAGC_RISE_SLOW);
- }
- }
- }
+ /* Normalize the proposed new mixer level */
+ if ( new_mixer_level < 0.05f )
+ new_mixer_level = 0.05f;
+ else if ( new_mixer_level > 1.00f )
+ new_mixer_level = 1.00f;
+
+ if ( new_mixer_level != mixer_level )
+ iaxc_input_level_set(new_mixer_level);
}
- /* This is ugly. Basically just don't get volume level if speex thought
- * we were silent. Just set it to 0 in that case */
- if ( iaxci_silence_threshold > 0.0f && silent )
- input_level = 0.0f;
-
do_level_callback();
- volume = vol_to_db(input_level);
-
- if ( volume < lowest_volume )
- lowest_volume = volume;
-
- if ( iaxci_silence_threshold > 0.0f )
- return silent;
- else
- return volume < iaxci_silence_threshold;
+ return using_vad ? !speaking :
+ vol_to_db(input_level) < iaxci_silence_threshold;
}
static int output_postprocess(const short * audio, int len)
@@ -517,6 +556,36 @@
rb_ReadRingBuffer(&ecOutRing, delayedBuf, SAMPLES_PER_FRAME * 2);
+ /* TODO: speex_echo_cancellation() and speex_preprocess_run() operate
+ * on the same state and thus must be serialized. Because the audio
+ * lock is not held, this call has the potential to mess-up the
+ * preprocessor (which is serialized by the audio lock). I believe the
+ * net effect of this problem is to break residual echo cancellation
+ * when these calls overlap. Unfortunately, just serializing this
+ * speex_echo_cancellation() call with the audio lock may not be
+ * sufficient since the next call to speex_preprocess_run() is counting
+ * on operating on this cancelledBuffer -- since we buffer the input
+ * audio (cancelledBuffer), we are actually explicitly decoupling the
+ * calls to speex_echo_cancellation() and speex_preprocess_run(). Oops.
+ *
+ * In other words, it should go like this:
+ *
+ * speex_echo_cancellation(A)
+ * speex_preprocess_run(A)
+ * speex_echo_cancellation(B)
+ * speex_preprocess_run(B)
+ * speex_echo_cancellation(C)
+ * speex_preprocess_run(C)
+ *
+ * but it actually may be going like this:
+ *
+ * speex_echo_cancellation(A)
+ * speex_echo_cancellation(B)
+ * speex_preprocess_run(A) -- bad, residual echo from B is applied to A
+ * speex_echo_cancellation(C)
+ * speex_preprocess_run(B) -- bad, residual echo from C is applied to B
+ * speex_preprocess_run(C)
+ */
speex_echo_cancellation(ec, inputBuffer, delayedBuf, cancelledBuffer);
memcpy(inputBuffer, cancelledBuffer, samples * sizeof(short));
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
×
Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.
Thanks for helping keep SourceForge clean.
X