-void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
-{
- vector<float> samples_card;
- vector<float> samples_out;
-
- // TODO: Allow mixing audio from several sources.
- unsigned selected_audio_card = theme->map_signal(audio_source_channel);
- assert(selected_audio_card < num_cards);
-
- for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
- samples_card.resize(num_samples * 2);
- {
- unique_lock<mutex> lock(cards[card_index].audio_mutex);
- cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples);
- }
- if (card_index == selected_audio_card) {
- samples_out = move(samples_card);
- }
- }
-
- // Cut away everything under 120 Hz (or whatever the cutoff is);
- // we don't need it for voice, and it will reduce headroom
- // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
- // should be dampened.)
- if (locut_enabled) {
- locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
- }
-
- // Apply a level compressor to get the general level right.
- // Basically, if it's over about -40 dBFS, we squeeze it down to that level
- // (or more precisely, near it, since we don't use infinite ratio),
- // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
- // 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(compressor_mutex);
- if (level_compressor_enabled) {
- float threshold = 0.01f; // -40 dBFS.
- float ratio = 20.0f;
- float attack_time = 0.5f;
- float release_time = 20.0f;
- float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f); // +26 dB.
- level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
- gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
- } else {
- // Just apply the gain we already had.
- float g = pow(10.0f, gain_staging_db / 20.0f);
- for (size_t i = 0; i < samples_out.size(); ++i) {
- samples_out[i] *= g;
- }
- }
- }
-
-#if 0
- printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
- level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
- level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
- 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
-#endif
-
-// float limiter_att, compressor_att;
-
- // The real compressor.
- if (compressor_enabled) {
- float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
- float ratio = 20.0f;
- float attack_time = 0.005f;
- float release_time = 0.040f;
- float makeup_gain = 2.0f; // +6 dB.
- compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
-// compressor_att = compressor.get_attenuation();
- }
-
- // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
- // Note that since ratio is not infinite, we could go slightly higher than this.
- if (limiter_enabled) {
- float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
- float ratio = 30.0f;
- float attack_time = 0.0f; // Instant.
- float release_time = 0.020f;
- float makeup_gain = 1.0f; // 0 dB.
- limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
-// limiter_att = limiter.get_attenuation();
- }
-
-// printf("limiter=%+5.1f compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
-
- // Upsample 4x to find interpolated peak.
- peak_resampler.inp_data = samples_out.data();
- peak_resampler.inp_count = samples_out.size() / 2;
-
- vector<float> interpolated_samples_out;
- interpolated_samples_out.resize(samples_out.size());
- while (peak_resampler.inp_count > 0) { // About four iterations.
- peak_resampler.out_data = &interpolated_samples_out[0];
- peak_resampler.out_count = interpolated_samples_out.size() / 2;
- peak_resampler.process();
- size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
- peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
- peak_resampler.out_data = nullptr;
- }
-
- // 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);