Changeset 2199 for pjproject/trunk/pjmedia/src/pjmedia/echo_suppress.c

Timestamp:

Aug 10, 2008 4:15:14 PM (16 years ago)

Author:

bennylp

Message:

Ticket #590: new echo suppressor which should work much better than the old one

File:

• pjproject/trunk/pjmedia/src/pjmedia/echo_suppress.c (modified) (5 diffs)

pjproject/trunk/pjmedia/src/pjmedia/echo_suppress.c

@@ -18 +18 @@
  */
 #include <pjmedia/types.h>
+#include <pjmedia/alaw_ulaw.h>
 #include <pjmedia/errno.h>
 #include <pjmedia/silencedet.h>
+#include <pj/array.h>
 #include <pj/assert.h>
 #include <pj/lock.h>
@@ -30 +32 @@
 #define THIS_FILE                           "echo_suppress.c"
 
+/* Maximum float constant */
+#define MAX_FLOAT               (float)1.701411e38
+
+/* The effective learn duration (in seconds) before we declare that learning
+ * is complete. The actual learning duration itself may be longer depending
+ * on the conversation pattern (e.g. we can't detect echo if speaker is only
+ * playing silence).
+ */
+#define MAX_CALC_DURATION_SEC   3
+
+/* The internal audio segment length, in milliseconds. 10ms shold be good
+ * and no need to change it.
+ */
+#define SEGMENT_PTIME           10
+
+/* The length of the template signal in milliseconds. The longer the template,
+ * the better correlation will be found, at the expense of more processing
+ * and longer learning time.
+ */
+#define TEMPLATE_PTIME          200
+
+/* How long to look back in the past to see if either mic or speaker is
+ * active.
+ */
+#define SIGNAL_LOOKUP_MSEC      200
+
+/* The minimum level value to be considered as talking, in uLaw complement
+ * (0-255).
+ */
+#define MIN_SIGNAL_ULAW         35
+
+/* The period (in seconds) on which the ES will analize it's effectiveness,
+ * and it may trigger soft-reset to force recalculation.
+ */
+#define CHECK_PERIOD            30
+
+/* Maximum signal level of average echo residue (in uLaw complement). When
+ * the residue value exceeds this value, we force the ES to re-learn.
+ */
+#define MAX_RESIDUE             2.5
+
+
+#if 0
+#   define TRACE_(expr) PJ_LOG(5,expr)
+#else
+#   define TRACE_(expr)
+#endif
+
+PJ_INLINE(float) fabs(float val)
+{
+    if (val < 0)
+        return -val;
+    else
+        return val;
+}
+
+
+#if defined(PJ_HAS_FLOATING_POINT) && PJ_HAS_FLOATING_POINT!=0
+    typedef float pj_ufloat_t;
+#   define pj_ufloat_from_float(f)      (f)
+#   define pj_ufloat_mul_u(val1, f)     ((val1) * (f))
+#   define pj_ufloat_mul_i(val1, f)     ((val1) * (f))
+#else
+    typedef pj_uint32_t pj_ufloat_t;
+
+    pj_ufloat_t pj_ufloat_from_float(float f)
+    {
+        return (pj_ufloat_t)(f * 65536);
+    }
+
+    unsigned pj_ufloat_mul_u(unsigned val1, pj_ufloat_t val2)
+    {
+        return (val1 * val2) >> 16;
+    }
+
+    int pj_ufloat_mul_i(int val1, pj_ufloat_t val2)
+    {
+        return (val1 * (pj_int32_t)val2) >> 16;
+    }
+#endif
+
+
+/* Conversation state */
+typedef enum talk_state
+{
+    ST_NULL,
+    ST_LOCAL_TALK,
+    ST_REM_SILENT,
+    ST_DOUBLETALK,
+    ST_REM_TALK
+} talk_state;
+
+const char *state_names[] = 
+{
+    "Null",
+    "local talking",
+    "remote silent",
+    "doubletalk",
+    "remote talking"
+};
+
+
+/* Description:
+
+   The echo suppressor tries to find the position of echoed signal by looking
+   at the correlation between signal played to the speaker (played signal) 
+   and the signal captured from the microphone (recorded signal).
+
+   To do this, it first divides the frames (from mic and speaker) into 
+   segments, calculate the audio level of the segment, and save the level
+   information in the playback and record history (play_hist and rec_hist
+   respectively).
+
+   In the history, the newest element (depicted as "t0" in the diagram belo)
+   is put in the last position of the array.
+
+   The record history size is as large as the template size (tmpl_cnt), since
+   we will use the record history as the template to find the best matching 
+   position in the playback history.
+
+   Here is the record history buffer:
+
+       <--templ_cnt-->
+       +-------------+
+       |   rec_hist  |
+       +-------------+
+    t-templ_cnt......t0
+
+   As you can see, the newest frame ("t0") is put as the last element in the
+   array.
+
+   The playback history size is larger than record history, since we need to
+   find the matching pattern in the past. The playback history size is
+   "templ_cnt + tail_cnt", where "tail_cnt" is the number of segments equal
+   to the maximum tail length. The maximum tail length is set when the ES
+   is created.
+
+   Here is the playback history buffer:
+
+       <-----tail_cnt-----> <--templ_cnt-->
+       +-------------------+--------------+
+       |             play_hist            |
+       +-------------------+--------------+
+   t-play_hist_cnt...t-templ_cnt.......t0
+
+
+
+   Learning:
+
+   During the processing, the ES calculates the following values:
+    - the correlation value, that is how similar the playback signal compared
+      to the mic signal. The lower the correlation value the better (i.e. more
+      similar) the signal is. The correlation value is done over the template
+      duration.
+    - the gain scaling factor, that is the ratio between mic signal and 
+      speaker signal. The ES calculates both the minimum and average ratios.
+
+   The ES calculates both the values above for every tail position in the
+   playback history. The values are saved in arrays below:
+
+     <-----tail_cnt----->
+     +-------------------+
+     |      corr_sum     |
+     +-------------------+
+     |     min_factor    |
+     +-------------------+
+     |     avg_factor    |
+     +-------------------+
+
+   At the end of processing, the ES iterates through the correlation array and
+   picks the tail index with the lowest corr_sum value. This is the position
+   where echo is most likely to be found.
+
+
+   Processing:
+
+   Once learning is done, the ES will change the level of the mic signal 
+   depending on the state of the conversation and according to the ratio that
+   has been found in the learning phase above.
+
+ */
 
 /*
- * Simple echo suppresor
+ * The simple echo suppresor state
  */
 typedef struct echo_supp
 {
-    pjmedia_silence_det *sd;
-    unsigned             samples_per_frame;
-    unsigned             tail_ms;
+    unsigned     clock_rate;        /* Clock rate.                          */
+    pj_uint16_t  samples_per_frame; /* Frame length in samples              */
+    pj_uint16_t  samples_per_segment;/* Segment length in samples           */
+    pj_uint16_t  tail_ms;           /* Tail length in milliseconds          */
+    pj_uint16_t  tail_samples;      /* Tail length in samples.              */
+
+    pj_bool_t    learning;          /* Are we still learning yet?           */
+    talk_state   talk_state;        /* Current talking state                */
+    int          tail_index;        /* Echo location, -1 if not found       */
+
+    unsigned     max_calc;          /* # of calc before learning complete.
+                                       (see MAX_CALC_DURATION_SEC)          */
+    unsigned     calc_cnt;          /* Number of calculations so far        */
+
+    unsigned     update_cnt;        /* # of updates                         */
+    unsigned     templ_cnt;         /* Template length, in # of segments    */
+    unsigned     tail_cnt;          /* Tail length, in # of segments        */
+    unsigned     play_hist_cnt;     /* # of segments in play_hist           */
+    pj_uint16_t *play_hist;         /* Array of playback levels             */
+    pj_uint16_t *rec_hist;          /* Array of rec levels                  */
+
+    float       *corr_sum;          /* Array of corr for each tail pos.     */
+    float       *tmp_corr;          /* Temporary corr array calculation     */
+    float        best_corr;         /* Best correlation so far.             */
+
+    float       *min_factor;        /* Array of minimum scaling factor      */
+    float       *avg_factor;        /* Array of average scaling factor      */
+    float       *tmp_factor;        /* Array to store provisional result    */
+
+    unsigned     running_cnt;       /* Running duration in # of frames      */
+    float        residue;           /* Accummulated echo residue.           */
+    float        last_factor;       /* Last factor applied to mic signal    */
 } echo_supp;
 
@@ -55 +267 @@
 {
     echo_supp *ec;
-    pj_status_t status;
-
-    PJ_UNUSED_ARG(clock_rate);
+
     PJ_UNUSED_ARG(channel_count);
     PJ_UNUSED_ARG(options);
 
+    PJ_ASSERT_RETURN(samples_per_frame >= SEGMENT_PTIME * clock_rate / 1000,
+                     PJ_ENOTSUP);
+
     ec = PJ_POOL_ZALLOC_T(pool, struct echo_supp);
-    ec->samples_per_frame = samples_per_frame;
-    ec->tail_ms = tail_ms;
-
-    status = pjmedia_silence_det_create(pool, clock_rate, samples_per_frame,
-                                        &ec->sd);
-    if (status != PJ_SUCCESS)
-        return status;
-
-    pjmedia_silence_det_set_name(ec->sd, "ecsu%p");
-    pjmedia_silence_det_set_adaptive(ec->sd, PJMEDIA_ECHO_SUPPRESS_THRESHOLD);
-    pjmedia_silence_det_set_params(ec->sd, 100, 500, 3000);
+    ec->clock_rate = clock_rate;
+    ec->samples_per_frame = (pj_uint16_t)samples_per_frame;
+    ec->samples_per_segment = (pj_uint16_t)(SEGMENT_PTIME * clock_rate / 1000);
+    ec->tail_ms = (pj_uint16_t)tail_ms;
+    ec->tail_samples = (pj_uint16_t)(tail_ms * clock_rate / 1000);
+
+    ec->templ_cnt = TEMPLATE_PTIME / SEGMENT_PTIME;
+    ec->tail_cnt = (pj_uint16_t)(tail_ms / SEGMENT_PTIME);
+    ec->play_hist_cnt = (pj_uint16_t)(ec->tail_cnt+ec->templ_cnt);
+
+    ec->max_calc = (pj_uint16_t)(MAX_CALC_DURATION_SEC * clock_rate / 
+                                 ec->samples_per_segment);
+
+    ec->rec_hist = (pj_uint16_t*) 
+                    pj_pool_alloc(pool, ec->templ_cnt *
+                                        sizeof(ec->rec_hist[0]));
+
+    /* Note: play history has twice number of elements */
+    ec->play_hist = (pj_uint16_t*) 
+                     pj_pool_alloc(pool, ec->play_hist_cnt *
+                                         sizeof(ec->play_hist[0]));
+
+    ec->corr_sum = (float*)
+                   pj_pool_alloc(pool, ec->tail_cnt * 
+                                       sizeof(ec->corr_sum[0]));
+    ec->tmp_corr = (float*)
+                   pj_pool_alloc(pool, ec->tail_cnt * 
+                                       sizeof(ec->tmp_corr[0]));
+    ec->min_factor = (float*)
+                     pj_pool_alloc(pool, ec->tail_cnt * 
+                                         sizeof(ec->min_factor[0]));
+    ec->avg_factor = (float*)
+                     pj_pool_alloc(pool, ec->tail_cnt * 
+                                         sizeof(ec->avg_factor[0]));
+    ec->tmp_factor = (float*)
+                     pj_pool_alloc(pool, ec->tail_cnt * 
+                                         sizeof(ec->tmp_factor[0]));
+    echo_supp_reset(ec);
 
     *p_state = ec;
@@ -90 +330 @@
 
 /*
- * Reset
+ * Hard reset
  */
 PJ_DEF(void) echo_supp_reset(void *state)
 {
-    PJ_UNUSED_ARG(state);
-    return;
+    unsigned i;
+    echo_supp *ec = (echo_supp*) state;
+
+    pj_bzero(ec->rec_hist, ec->templ_cnt * sizeof(ec->rec_hist[0]));
+    pj_bzero(ec->play_hist, ec->play_hist_cnt * sizeof(ec->play_hist[0]));
+
+    for (i=0; i<ec->tail_cnt; ++i) {
+        ec->corr_sum[i] = ec->avg_factor[i] = 0;
+        ec->min_factor[i] = MAX_FLOAT;
+    }
+
+    ec->update_cnt = 0;
+    ec->calc_cnt = 0;
+    ec->learning = PJ_TRUE;
+    ec->tail_index = -1;
+    ec->best_corr = MAX_FLOAT;
+    ec->talk_state = ST_NULL;
+    ec->last_factor = 1.0;
+    ec->residue = 0;
+    ec->running_cnt = 0;
 }
 
 /*
+ * Soft reset to force the EC to re-learn without having to discard all
+ * rec and playback history.
+ */
+PJ_DEF(void) echo_supp_soft_reset(void *state)
+{
+    unsigned i;
+
+    echo_supp *ec = (echo_supp*) state;
+
+    for (i=0; i<ec->tail_cnt; ++i) {
+        ec->corr_sum[i] = 0;
+    }
+
+    ec->update_cnt = 0;
+    ec->calc_cnt = 0;
+    ec->learning = PJ_TRUE;
+    ec->best_corr = MAX_FLOAT;
+    ec->residue = 0;
+    ec->running_cnt = 0;
+
+    PJ_LOG(4,(THIS_FILE, "Echo suppressor soft reset. Re-learning.."));
+}
+
+
+/* Set state */
+static void echo_supp_set_state(echo_supp *ec, enum talk_state state)
+{
+    if (state != ec->talk_state) {
+        TRACE_((THIS_FILE, "[%03d.%03d] %s --> %s", 
+                           (ec->update_cnt * SEGMENT_PTIME / 1000), 
+                           ((ec->update_cnt * SEGMENT_PTIME) % 1000),
+                           state_names[ec->talk_state],
+                           state_names[state]));
+        ec->talk_state = state;
+    }
+}
+
+/*
+ * Update EC state
+ */
+static void echo_supp_update(echo_supp *ec, pj_int16_t *rec_frm,
+                             const pj_int16_t *play_frm)
+{
+    int prev_index;
+    unsigned i, frm_level, sum_rec_level;
+    float rec_corr;
+
+    ++ec->update_cnt;
+    if (ec->update_cnt > 0x7FFFFFFF)
+        ec->update_cnt = 0x7FFFFFFF; /* Detect overflow */
+
+    /* Calculate current play frame level */
+    frm_level = pjmedia_calc_avg_signal(play_frm, ec->samples_per_segment);
+    ++frm_level; /* to avoid division by zero */
+
+    /* Push current frame level to the back of the play history */
+    pj_array_erase(ec->play_hist, sizeof(pj_uint16_t), ec->play_hist_cnt, 0);
+    ec->play_hist[ec->play_hist_cnt-1] = (pj_uint16_t) frm_level;
+
+    /* Calculate level of current mic frame */
+    frm_level = pjmedia_calc_avg_signal(rec_frm, ec->samples_per_segment);
+    ++frm_level; /* to avoid division by zero */
+
+    /* Push to the back of the rec history */
+    pj_array_erase(ec->rec_hist, sizeof(pj_uint16_t), ec->templ_cnt, 0);
+    ec->rec_hist[ec->templ_cnt-1] = (pj_uint16_t) frm_level;
+
+
+    /* Can't do the calc until the play history is full. */
+    if (ec->update_cnt < ec->play_hist_cnt)
+        return;
+
+    /* Skip if learning is done */
+    if (!ec->learning)
+        return;
+
+
+    /* Calculate rec signal pattern */
+    rec_corr = 0;
+    sum_rec_level = 0;
+    for (i=0; i < ec->templ_cnt-1; ++i) {
+        float corr;
+        corr = (float)ec->rec_hist[i+1] / ec->rec_hist[i];
+        rec_corr += corr;
+        sum_rec_level += ec->rec_hist[i];
+    }
+    sum_rec_level += ec->rec_hist[i];
+
+    /* Iterate through the play history and calculate the signal correlation
+     * for every tail position in the play_hist. Save the result in temporary
+     * array since we may bail out early if the conversation state is not good
+     * to detect echo.
+     */
+    for (i=0; i < ec->tail_cnt; ++i) {
+        unsigned j, end, sum_play_level, ulaw;
+        float play_corr = 0, corr_diff;
+
+        sum_play_level = 0;
+        for (j=i, end=i+ec->templ_cnt-1; j<end; ++j) {
+            float corr;
+            corr = (float)ec->play_hist[j+1] / ec->play_hist[j];
+            play_corr += corr;
+            sum_play_level += ec->play_hist[j];
+        }
+        sum_play_level += ec->play_hist[j];
+
+        /* Bail out if remote isn't talking */
+        ulaw = pjmedia_linear2ulaw(sum_play_level/ec->templ_cnt) ^ 0xFF;
+        if (ulaw < MIN_SIGNAL_ULAW) {
+            echo_supp_set_state(ec, ST_REM_SILENT);
+            return;
+        }
+
+        /* Bail out if local user is talking */
+        if (sum_rec_level >= sum_play_level) {
+            echo_supp_set_state(ec, ST_LOCAL_TALK);
+            return;
+        }
+
+        /* Also bail out if we suspect there's a doubletalk */
+        ulaw = pjmedia_linear2ulaw(sum_rec_level/ec->templ_cnt) ^ 0xFF;
+        if (ulaw > MIN_SIGNAL_ULAW) {
+            echo_supp_set_state(ec, ST_DOUBLETALK);
+            return;
+        }
+
+        /* Calculate correlation and save to temporary array */
+        corr_diff = fabs(play_corr - rec_corr);
+        ec->tmp_corr[i] = corr_diff;
+
+        /* Also calculate the gain factor between mic and speaker level */
+        ec->tmp_factor[i] = (float)sum_rec_level / sum_play_level;
+        pj_assert(ec->tmp_factor[i] < 1);
+    }
+
+    /* We seem to have good signal, we can update the EC state */
+    echo_supp_set_state(ec, ST_REM_TALK);
+
+    /* Accummulate the correlation value to the history and at the same
+     * time find the tail index of the best correlation.
+     */
+    prev_index = ec->tail_index;
+    for (i=1; i<ec->tail_cnt-1; ++i) {
+        float *p = &ec->corr_sum[i], sum;
+
+        /* Accummulate correlation value  for this tail position */
+        ec->corr_sum[i] += ec->tmp_corr[i];
+
+        /* Update the min and avg gain factor for this tail position */
+        if (ec->tmp_factor[i] < ec->min_factor[i])
+            ec->min_factor[i] = ec->tmp_factor[i];
+        ec->avg_factor[i] = ((ec->avg_factor[i] * ec->tail_cnt) + 
+                                    ec->tmp_factor[i]) /
+                            (ec->tail_cnt + 1);
+
+        /* To get the best correlation, also include the correlation
+         * value of the neighbouring tail locations.
+         */
+        sum = *(p-1) + (*p)*2 + *(p+1);
+        //sum = *p;
+
+        /* See if we have better correlation value */
+        if (sum < ec->best_corr) {
+            ec->tail_index = i;
+            ec->best_corr = sum;
+        }
+    }
+
+    if (ec->tail_index != prev_index) {
+        unsigned duration;
+        int imin, iavg;
+
+        duration = ec->update_cnt * SEGMENT_PTIME;
+        imin = (int)(ec->min_factor[ec->tail_index] * 1000);
+        iavg = (int)(ec->avg_factor[ec->tail_index] * 1000);
+
+        PJ_LOG(4,(THIS_FILE, 
+                  "Echo suppressor updated at t=%03d.%03ds, echo tail=%d msec"
+                  ", factor min/avg=%d.%03d/%d.%03d",
+                  (duration/1000), (duration%1000),
+                  (ec->tail_cnt-ec->tail_index) * SEGMENT_PTIME,
+                  imin/1000, imin%1000,
+                  iavg/1000, iavg%1000));
+
+    }
+
+    ++ec->calc_cnt;
+
+    if (ec->calc_cnt > ec->max_calc) {
+        unsigned duration;
+        int imin, iavg;
+
+
+        ec->learning = PJ_FALSE;
+        ec->running_cnt = 0;
+
+        duration = ec->update_cnt * SEGMENT_PTIME;
+        imin = (int)(ec->min_factor[ec->tail_index] * 1000);
+        iavg = (int)(ec->avg_factor[ec->tail_index] * 1000);
+
+        PJ_LOG(4,(THIS_FILE, 
+                  "Echo suppressor learning done at t=%03d.%03ds, tail=%d ms"
+                  ", factor min/avg=%d.%03d/%d.%03d",
+                  (duration/1000), (duration%1000),
+                  (ec->tail_cnt-ec->tail_index) * SEGMENT_PTIME,
+                  imin/1000, imin%1000,
+                  iavg/1000, iavg%1000));
+    }
+
+}
+
+
+/* Amplify frame */
+static void amplify_frame(pj_int16_t *frm, unsigned length, 
+                          pj_ufloat_t factor)
+{
+    unsigned i;
+
+    for (i=0; i<length; ++i) {
+        frm[i] = (pj_int16_t)pj_ufloat_mul_i(frm[i], factor);
+    }
+}
+
+/* 
  * Perform echo cancellation.
  */
@@ -107 +589 @@
                                            void *reserved )
 {
+    unsigned i, N;
     echo_supp *ec = (echo_supp*) state;
-    pj_bool_t silence;
 
     PJ_UNUSED_ARG(options);
     PJ_UNUSED_ARG(reserved);
 
-    silence = pjmedia_silence_det_detect(ec->sd, play_frm, 
-                                         ec->samples_per_frame, NULL);
-
-    if (!silence) {
-#if defined(PJMEDIA_ECHO_SUPPRESS_FACTOR) && PJMEDIA_ECHO_SUPPRESS_FACTOR!=0
-        unsigned i;
-        for (i=0; i<ec->samples_per_frame; ++i) {
-            rec_frm[i] = (pj_int16_t)(rec_frm[i] >> 
-                                      PJMEDIA_ECHO_SUPPRESS_FACTOR);
-        }
-#else
-        pjmedia_zero_samples(rec_frm, ec->samples_per_frame);
-#endif
+    /* Calculate number of segments. This should be okay even if
+     * samples_per_frame is not a multiply of samples_per_segment, since
+     * we only calculate level.
+     */
+    N = ec->samples_per_frame / ec->samples_per_segment;
+    pj_assert(N>0);
+    for (i=0; i<N; ++i) {
+        unsigned pos = i * ec->samples_per_segment;
+        echo_supp_update(ec, rec_frm+pos, play_frm+pos);
+    }
+
+    if (ec->tail_index < 0) {
+        /* Not ready */
+    } else {
+        unsigned lookup_cnt, rec_level=0, play_level=0;
+        unsigned tail_cnt;
+        float factor;
+
+        /* How many previous segments to lookup */
+        lookup_cnt = SIGNAL_LOOKUP_MSEC / SEGMENT_PTIME;
+        if (lookup_cnt > ec->templ_cnt)
+            lookup_cnt = ec->templ_cnt;
+
+        /* Lookup in recording history to get maximum mic level, to see
+         * if local user is currently talking
+         */
+        for (i=ec->templ_cnt - lookup_cnt; i < ec->templ_cnt; ++i) {
+            if (ec->rec_hist[i] > rec_level)
+                rec_level = ec->rec_hist[i];
+        }
+        rec_level = pjmedia_linear2ulaw(rec_level) ^ 0xFF;
+
+        /* Calculate the detected tail length, in # of segments */
+        tail_cnt = (ec->tail_cnt - ec->tail_index);
+
+        /* Lookup in playback history to get max speaker level, to see
+         * if remote user is currently talking
+         */
+        for (i=ec->play_hist_cnt -lookup_cnt -tail_cnt; 
+             i<ec->play_hist_cnt-tail_cnt; ++i) 
+        {
+            if (ec->play_hist[i] > play_level)
+                play_level = ec->play_hist[i];
+        }
+        play_level = pjmedia_linear2ulaw(play_level) ^ 0xFF;
+
+        if (rec_level >= MIN_SIGNAL_ULAW) {
+            if (play_level < MIN_SIGNAL_ULAW) {
+                /* Mic is talking, speaker is idle. Let mic signal pass as is.
+                 */
+                factor = 1.0;
+                echo_supp_set_state(ec, ST_LOCAL_TALK);
+            } else {
+                /* Seems that both are talking. Scale the mic signal
+                 * down a little bit to reduce echo, while allowing both
+                 * parties to talk at the same time.
+                 */
+                factor = (float)(ec->avg_factor[ec->tail_index] * 2);
+                echo_supp_set_state(ec, ST_DOUBLETALK);
+            }
+        } else {
+            if (play_level < MIN_SIGNAL_ULAW) {
+                /* Both mic and speaker seems to be idle. Also scale the
+                 * mic signal down with average factor to reduce low power
+                 * echo.
+                 */
+                factor = ec->avg_factor[ec->tail_index] * 3 / 2;
+                echo_supp_set_state(ec, ST_REM_SILENT);
+            } else {
+                /* Mic is idle, but there's something playing in speaker.
+                 * Scale the mic down to minimum
+                 */
+                factor = ec->min_factor[ec->tail_index] / 2;
+                echo_supp_set_state(ec, ST_REM_TALK);
+            }
+        }
+
+        /* Smoothen the transition */
+        if (factor > ec->last_factor)
+            factor = (factor + ec->last_factor) / 2;
+        else
+            factor = (factor + ec->last_factor*9) / 10;
+
+        /* Amplify frame */
+        amplify_frame(rec_frm, ec->samples_per_frame, 
+                      pj_ufloat_from_float(factor));
+        ec->last_factor = factor;
+
+        if (ec->talk_state == ST_REM_TALK) {
+            unsigned level, recalc_cnt;
+
+            /* Get the adjusted frame signal level */
+            level = pjmedia_calc_avg_signal(rec_frm, ec->samples_per_frame);
+            level = pjmedia_linear2ulaw(level) ^ 0xFF;
+
+            /* Accumulate average echo residue to see the ES effectiveness */
+            ec->residue = ((ec->residue * ec->running_cnt) + level) / 
+                          (ec->running_cnt + 1);
+
+            ++ec->running_cnt;
+
+            /* Check if we need to re-learn */
+            recalc_cnt = CHECK_PERIOD * ec->clock_rate / ec->samples_per_frame;
+            if (ec->running_cnt > recalc_cnt) {
+                int iresidue;
+
+                iresidue = (int)(ec->residue*1000);
+
+                PJ_LOG(5,(THIS_FILE, "Echo suppressor residue = %d.%03d",
+                          iresidue/1000, iresidue%1000));
+
+                if (ec->residue > MAX_RESIDUE && !ec->learning) {
+                    echo_supp_soft_reset(ec);
+                    ec->residue = 0;
+                } else {
+                    ec->running_cnt = 0;
+                    ec->residue = 0;
+                }
+            }
+        }
     }