+/**
+ * @name Postfilter functions
+ * Postfilter functions (gain control, wiener denoise filter, DC filter,
+ * kalman smoothening, plus surrounding code to wrap it)
+ * @{
+ */
+/**
+ * Adaptive gain control (as used in postfilter).
+ *
+ * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
+ * that the energy here is calculated using sum(abs(...)), whereas the
+ * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
+ *
+ * @param out output buffer for filtered samples
+ * @param in input buffer containing the samples as they are after the
+ * postfilter steps so far
+ * @param speech_synth input buffer containing speech synth before postfilter
+ * @param size input buffer size
+ * @param alpha exponential filter factor
+ * @param gain_mem pointer to filter memory (single float)
+ */
+static void adaptive_gain_control(float *out, const float *in,
+ const float *speech_synth,
+ int size, float alpha, float *gain_mem)
+{
+ int i;
+ float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
+ float mem = *gain_mem;
+
+ for (i = 0; i < size; i++) {
+ speech_energy += fabsf(speech_synth[i]);
+ postfilter_energy += fabsf(in[i]);
+ }
+ gain_scale_factor = (1.0 - alpha) * speech_energy / postfilter_energy;
+
+ for (i = 0; i < size; i++) {
+ mem = alpha * mem + gain_scale_factor;
+ out[i] = in[i] * mem;
+ }
+
+ *gain_mem = mem;
+}
+
+/**
+ * Kalman smoothing function.
+ *
+ * This function looks back pitch +/- 3 samples back into history to find
+ * the best fitting curve (that one giving the optimal gain of the two
+ * signals, i.e. the highest dot product between the two), and then
+ * uses that signal history to smoothen the output of the speech synthesis
+ * filter.
+ *
+ * @param s WMA Voice decoding context
+ * @param pitch pitch of the speech signal
+ * @param in input speech signal
+ * @param out output pointer for smoothened signal
+ * @param size input/output buffer size
+ *
+ * @returns -1 if no smoothening took place, e.g. because no optimal
+ * fit could be found, or 0 on success.
+ */
+static int kalman_smoothen(WMAVoiceContext *s, int pitch,
+ const float *in, float *out, int size)
+{
+ int n;
+ float optimal_gain = 0, dot;
+ const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],
+ *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
+ *best_hist_ptr;
+
+ /* find best fitting point in history */
+ do {
+ dot = ff_dot_productf(in, ptr, size);
+ if (dot > optimal_gain) {
+ optimal_gain = dot;
+ best_hist_ptr = ptr;
+ }
+ } while (--ptr >= end);
+
+ if (optimal_gain <= 0)
+ return -1;
+ dot = ff_dot_productf(best_hist_ptr, best_hist_ptr, size);
+ if (dot <= 0) // would be 1.0
+ return -1;
+
+ if (optimal_gain <= dot) {
+ dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
+ } else
+ dot = 0.625;
+
+ /* actual smoothing */
+ for (n = 0; n < size; n++)
+ out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
+
+ return 0;
+}
+
+/**
+ * Get the tilt factor of a formant filter from its transfer function
+ * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
+ * but somehow (??) it does a speech synthesis filter in the
+ * middle, which is missing here
+ *
+ * @param lpcs LPC coefficients
+ * @param n_lpcs Size of LPC buffer
+ * @returns the tilt factor
+ */
+static float tilt_factor(const float *lpcs, int n_lpcs)
+{
+ float rh0, rh1;
+
+ rh0 = 1.0 + ff_dot_productf(lpcs, lpcs, n_lpcs);
+ rh1 = lpcs[0] + ff_dot_productf(lpcs, &lpcs[1], n_lpcs - 1);
+
+ return rh1 / rh0;
+}
+
+/**
+ * Derive denoise filter coefficients (in real domain) from the LPCs.
+ */
+static void calc_input_response(WMAVoiceContext *s, float *lpcs,
+ int fcb_type, float *coeffs, int remainder)
+{
+ float last_coeff, min = 15.0, max = -15.0;
+ float irange, angle_mul, gain_mul, range, sq;
+ int n, idx;
+
+ /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
+ s->rdft.rdft_calc(&s->rdft, lpcs);
+#define log_range(var, assign) do { \
+ float tmp = log10f(assign); var = tmp; \
+ max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
+ } while (0)
+ log_range(last_coeff, lpcs[1] * lpcs[1]);
+ for (n = 1; n < 64; n++)
+ log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] +
+ lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
+ log_range(lpcs[0], lpcs[0] * lpcs[0]);
+#undef log_range
+ range = max - min;
+ lpcs[64] = last_coeff;
+
+ /* Now, use this spectrum to pick out these frequencies with higher
+ * (relative) power/energy (which we then take to be "not noise"),
+ * and set up a table (still in lpc[]) of (relative) gains per frequency.
+ * These frequencies will be maintained, while others ("noise") will be
+ * decreased in the filter output. */
+ irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
+ gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) :
+ (5.0 / 14.7));
+ angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);
+ for (n = 0; n <= 64; n++) {
+ float pwr;
+
+ idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1);
+ pwr = wmavoice_denoise_power_table[s->denoise_strength][idx];
+ lpcs[n] = angle_mul * pwr;
+
+ /* 70.57 =~ 1/log10(1.0331663) */
+ idx = (pwr * gain_mul - 0.0295) * 70.570526123;
+ if (idx > 127) { // fallback if index falls outside table range
+ coeffs[n] = wmavoice_energy_table[127] *
+ powf(1.0331663, idx - 127);
+ } else
+ coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
+ }
+
+ /* calculate the Hilbert transform of the gains, which we do (since this
+ * is a sinus input) by doing a phase shift (in theory, H(sin())=cos()).
+ * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
+ * "moment" of the LPCs in this filter. */
+ s->dct.dct_calc(&s->dct, lpcs);
+ s->dst.dct_calc(&s->dst, lpcs);
+
+ /* Split out the coefficient indexes into phase/magnitude pairs */
+ idx = 255 + av_clip(lpcs[64], -255, 255);
+ coeffs[0] = coeffs[0] * s->cos[idx];
+ idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
+ last_coeff = coeffs[64] * s->cos[idx];
+ for (n = 63;; n--) {
+ idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
+ coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
+ coeffs[n * 2] = coeffs[n] * s->cos[idx];
+
+ if (!--n) break;
+
+ idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
+ coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
+ coeffs[n * 2] = coeffs[n] * s->cos[idx];
+ }
+ coeffs[1] = last_coeff;
+
+ /* move into real domain */
+ s->irdft.rdft_calc(&s->irdft, coeffs);
+
+ /* tilt correction and normalize scale */
+ memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder));
+ if (s->denoise_tilt_corr) {
+ float tilt_mem = 0;
+
+ coeffs[remainder - 1] = 0;
+ ff_tilt_compensation(&tilt_mem,
+ -1.8 * tilt_factor(coeffs, remainder - 1),
+ coeffs, remainder);
+ }
+ sq = (1.0 / 64.0) * sqrtf(1 / ff_dot_productf(coeffs, coeffs, remainder));
+ for (n = 0; n < remainder; n++)
+ coeffs[n] *= sq;
+}
+
+/**
+ * This function applies a Wiener filter on the (noisy) speech signal as
+ * a means to denoise it.
+ *
+ * - take RDFT of LPCs to get the power spectrum of the noise + speech;
+ * - using this power spectrum, calculate (for each frequency) the Wiener
+ * filter gain, which depends on the frequency power and desired level
+ * of noise subtraction (when set too high, this leads to artifacts)
+ * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
+ * of 4-8kHz);
+ * - by doing a phase shift, calculate the Hilbert transform of this array
+ * of per-frequency filter-gains to get the filtering coefficients;
+ * - smoothen/normalize/de-tilt these filter coefficients as desired;
+ * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
+ * to get the denoised speech signal;
+ * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
+ * the frame boundary) are saved and applied to subsequent frames by an
+ * overlap-add method (otherwise you get clicking-artifacts).
+ *
+ * @param s WMA Voice decoding context
+ * @param fcb_type Frame (codebook) type
+ * @param synth_pf input: the noisy speech signal, output: denoised speech
+ * data; should be 16-byte aligned (for ASM purposes)
+ * @param size size of the speech data
+ * @param lpcs LPCs used to synthesize this frame's speech data
+ */
+static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
+ float *synth_pf, int size,
+ const float *lpcs)
+{
+ int remainder, lim, n;
+
+ if (fcb_type != FCB_TYPE_SILENCE) {
+ float *tilted_lpcs = s->tilted_lpcs_pf,
+ *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
+
+ tilted_lpcs[0] = 1.0;
+ memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
+ memset(&tilted_lpcs[s->lsps + 1], 0,
+ sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
+ ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
+ tilted_lpcs, s->lsps + 2);
+
+ /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
+ * size is applied to the next frame. All input beyond this is zero,
+ * and thus all output beyond this will go towards zero, hence we can
+ * limit to min(size-1, 127-size) as a performance consideration. */
+ remainder = FFMIN(127 - size, size - 1);
+ calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
+
+ /* apply coefficients (in frequency spectrum domain), i.e. complex
+ * number multiplication */
+ memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
+ s->rdft.rdft_calc(&s->rdft, synth_pf);
+ s->rdft.rdft_calc(&s->rdft, coeffs);
+ synth_pf[0] *= coeffs[0];
+ synth_pf[1] *= coeffs[1];
+ for (n = 1; n < 64; n++) {
+ float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];
+ synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1];
+ synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];
+ }
+ s->irdft.rdft_calc(&s->irdft, synth_pf);
+ }
+
+ /* merge filter output with the history of previous runs */
+ if (s->denoise_filter_cache_size) {
+ lim = FFMIN(s->denoise_filter_cache_size, size);
+ for (n = 0; n < lim; n++)
+ synth_pf[n] += s->denoise_filter_cache[n];
+ s->denoise_filter_cache_size -= lim;
+ memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
+ sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
+ }
+
+ /* move remainder of filter output into a cache for future runs */
+ if (fcb_type != FCB_TYPE_SILENCE) {
+ lim = FFMIN(remainder, s->denoise_filter_cache_size);
+ for (n = 0; n < lim; n++)
+ s->denoise_filter_cache[n] += synth_pf[size + n];
+ if (lim < remainder) {
+ memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
+ sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
+ s->denoise_filter_cache_size = remainder;
+ }
+ }
+}
+
+/**
+ * Averaging projection filter, the postfilter used in WMAVoice.
+ *
+ * This uses the following steps:
+ * - A zero-synthesis filter (generate excitation from synth signal)
+ * - Kalman smoothing on excitation, based on pitch
+ * - Re-synthesized smoothened output
+ * - Iterative Wiener denoise filter
+ * - Adaptive gain filter
+ * - DC filter
+ *
+ * @param s WMAVoice decoding context
+ * @param synth Speech synthesis output (before postfilter)
+ * @param samples Output buffer for filtered samples
+ * @param size Buffer size of synth & samples
+ * @param lpcs Generated LPCs used for speech synthesis
+ * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
+ * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
+ * @param pitch Pitch of the input signal
+ */
+static void postfilter(WMAVoiceContext *s, const float *synth,
+ float *samples, int size,
+ const float *lpcs, float *zero_exc_pf,
+ int fcb_type, int pitch)
+{
+ float synth_filter_in_buf[MAX_FRAMESIZE / 2],
+ *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
+ *synth_filter_in = zero_exc_pf;
+
+ assert(size <= MAX_FRAMESIZE / 2);
+
+ /* generate excitation from input signal */
+ ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
+
+ if (fcb_type >= FCB_TYPE_AW_PULSES &&
+ !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
+ synth_filter_in = synth_filter_in_buf;
+
+ /* re-synthesize speech after smoothening, and keep history */
+ ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
+ synth_filter_in, size, s->lsps);
+ memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
+ sizeof(synth_pf[0]) * s->lsps);
+
+ wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
+
+ adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
+ &s->postfilter_agc);
+
+ if (s->dc_level > 8) {
+ /* remove ultra-low frequency DC noise / highpass filter;
+ * coefficients are identical to those used in SIPR decoding,
+ * and very closely resemble those used in AMR-NB decoding. */
+ ff_acelp_apply_order_2_transfer_function(samples, samples,
+ (const float[2]) { -1.99997, 1.0 },
+ (const float[2]) { -1.9330735188, 0.93589198496 },
+ 0.93980580475, s->dcf_mem, size);
+ }
+}
+/**
+ * @}
+ */
+