Add a final makeup gain, trying to set the level straight at +0 LU (more or less...

[nageru] / mixer.cpp

diff --git a/mixer.cpp b/mixer.cpp
index 72022026f771056e0211ccac874e69c8ee0ed22a..f3ff4c385c1d46b4f3b03dc5832a6ba816a8587a 100644 (file)
--- a/mixer.cpp
+++ b/mixer.cpp
@@ -534,7 +534,7 @@ void Mixer::thread_func()
 
                        audio_level_callback(loudness_s, 20.0 * log10(peak),
                                             loudness_i, loudness_range_low, loudness_range_high,
-                                            gain_staging_db);
+                                            gain_staging_db, 20.0 * log10(final_makeup_gain));
                }
 
                for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
@@ -716,7 +716,7 @@ void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
        // entirely arbitrary, but from practical tests with speech, it seems to
        // put ut around -23 LUFS, so it's a reasonable starting point for later use.
        {
-               unique_lock<mutex> lock(level_compressor_mutex);
+               unique_lock<mutex> lock(compressor_mutex);
                if (level_compressor_enabled) {
                        float threshold = 0.01f;   // -40 dBFS.
                        float ratio = 20.0f;
@@ -782,6 +782,46 @@ void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
                peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
        }
 
+       // At this point, we are most likely close to +0 LU, but all of our
+       // measurements have been on raw sample values, not R128 values.
+       // So we have a final makeup gain to get us to +0 LU; the gain
+       // adjustments required should be relatively small, and also, the
+       // offset shouldn't change much (only if the type of audio changes
+       // significantly). Thus, we shoot for updating this value basically
+       // “whenever we process buffers”, since the R128 calculation isn't exactly
+       // something we get out per-sample.
+       //
+       // Note that there's a feedback loop here, so we choose a very slow filter
+       // (half-time of 100 seconds).
+       double target_loudness_factor, alpha;
+       {
+               unique_lock<mutex> lock(compressor_mutex);
+               double loudness_lu = r128.loudness_M() - ref_level_lufs;
+               double current_makeup_lu = 20.0f * log10(final_makeup_gain);
+               target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f);
+
+               // If we're outside +/- 5 LU uncorrected, we don't count it as
+               // a normal signal (probably silence) and don't change the
+               // correction factor; just apply what we already have.
+               if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) {
+                       alpha = 0.0;
+               } else {
+                       // Formula adapted from
+                       // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter.
+                       const double half_time_s = 100.0;
+                       const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY);
+                       alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0);
+               }
+
+               double m = final_makeup_gain;
+               for (size_t i = 0; i < samples_out.size(); i += 2) {
+                       samples_out[i + 0] *= m;
+                       samples_out[i + 1] *= m;
+                       m += (target_loudness_factor - m) * alpha;
+               }
+               final_makeup_gain = m;
+       }
+
        // Find R128 levels.
        vector<float> left, right;
        deinterleave_samples(samples_out, &left, &right);