* Windows Media Audio Voice decoder.
* Copyright (c) 2009 Ronald S. Bultje
*
- * This file is part of FFmpeg.
+ * This file is part of Libav.
*
- * FFmpeg is free software; you can redistribute it and/or
+ * Libav is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
- * FFmpeg is distributed in the hope that it will be useful,
+ * Libav is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
- * License along with FFmpeg; if not, write to the Free Software
+ * License along with Libav; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
* @author Ronald S. Bultje <rsbultje@gmail.com>
*/
+#define UNCHECKED_BITSTREAM_READER 1
+
#include <math.h>
+
+#include "libavutil/channel_layout.h"
+#include "libavutil/float_dsp.h"
+#include "libavutil/mem.h"
#include "avcodec.h"
+#include "internal.h"
#include "get_bits.h"
#include "put_bits.h"
#include "wmavoice_data.h"
-#include "celp_math.h"
#include "celp_filters.h"
#include "acelp_vectors.h"
#include "acelp_filters.h"
#include "lsp.h"
-#include "libavutil/lzo.h"
+#include "dct.h"
+#include "rdft.h"
+#include "sinewin.h"
#define MAX_BLOCKS 8 ///< maximum number of blocks per frame
#define MAX_LSPS 16 ///< maximum filter order
+#define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple
+ ///< of 16 for ASM input buffer alignment
#define MAX_FRAMES 3 ///< maximum number of frames per superframe
#define MAX_FRAMESIZE 160 ///< maximum number of samples per frame
#define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history
/**
* WMA Voice decoding context.
*/
-typedef struct {
+typedef struct WMAVoiceContext {
/**
- * @defgroup struct_global Global values
- * Global values, specified in the stream header / extradata or used
- * all over.
+ * @name Global values specified in the stream header / extradata or used all over.
* @{
*/
GetBitContext gb; ///< packet bitreader. During decoder init,
int history_nsamples; ///< number of samples in history for signal
///< prediction (through ACB)
+ /* postfilter specific values */
int do_apf; ///< whether to apply the averaged
///< projection filter (APF)
+ int denoise_strength; ///< strength of denoising in Wiener filter
+ ///< [0-11]
+ int denoise_tilt_corr; ///< Whether to apply tilt correction to the
+ ///< Wiener filter coefficients (postfilter)
+ int dc_level; ///< Predicted amount of DC noise, based
+ ///< on which a DC removal filter is used
int lsps; ///< number of LSPs per frame [10 or 16]
int lsp_q_mode; ///< defines quantizer defaults [0, 1]
/**
* @}
- * @defgroup struct_packet Packet values
- * Packet values, specified in the packet header or related to a packet.
+ *
+ * @name Packet values specified in the packet header or related to a packet.
+ *
* A packet is considered to be a single unit of data provided to this
* decoder by the demuxer.
* @{
*/
int spillover_nbits; ///< number of bits of the previous packet's
- ///< last superframe preceeding this
+ ///< last superframe preceding this
///< packet's first full superframe (useful
///< for re-synchronization also)
int has_residual_lsps; ///< if set, superframes contain one set of
/**
* @}
- * @defgroup struct_frame Frame and superframe values
+ *
+ * @name Frame and superframe values
* Superframe and frame data - these can change from frame to frame,
* although some of them do in that case serve as a cache / history for
* the next frame or superframe.
///< superframes, used as a history for
///< signal generation
float synth_history[MAX_LSPS]; ///< see #excitation_history
+ /**
+ * @}
+ *
+ * @name Postfilter values
+ *
+ * Variables used for postfilter implementation, mostly history for
+ * smoothing and so on, and context variables for FFT/iFFT.
+ * @{
+ */
+ RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the
+ ///< postfilter (for denoise filter)
+ DCTContext dct, dst; ///< contexts for phase shift (in Hilbert
+ ///< transform, part of postfilter)
+ float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi]
+ ///< range
+ float postfilter_agc; ///< gain control memory, used in
+ ///< #adaptive_gain_control()
+ float dcf_mem[2]; ///< DC filter history
+ float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];
+ ///< zero filter output (i.e. excitation)
+ ///< by postfilter
+ float denoise_filter_cache[MAX_FRAMESIZE];
+ int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
+ DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80];
+ ///< aligned buffer for LPC tilting
+ DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x80];
+ ///< aligned buffer for denoise coefficients
+ DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
+ ///< aligned buffer for postfilter speech
+ ///< synthesis
/**
* @}
*/
} WMAVoiceContext;
/**
- * Sets up the variable bit mode (VBM) tree from container extradata.
+ * Set up the variable bit mode (VBM) tree from container extradata.
* @param gb bit I/O context.
* The bit context (s->gb) should be loaded with byte 23-46 of the
* container extradata (i.e. the ones containing the VBM tree).
* @return 0 on success, <0 on error.
*/
static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
+{
+ int cntr[8] = { 0 }, n, res;
+
+ memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
+ for (n = 0; n < 17; n++) {
+ res = get_bits(gb, 3);
+ if (cntr[res] > 3) // should be >= 3 + (res == 7))
+ return -1;
+ vbm_tree[res * 3 + cntr[res]++] = n;
+ }
+ return 0;
+}
+
+static av_cold void wmavoice_init_static_data(AVCodec *codec)
{
static const uint8_t bits[] = {
2, 2, 2, 4, 4, 4,
0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10
0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx
};
- int cntr[8], n, res;
- memset(vbm_tree, 0xff, sizeof(vbm_tree));
- memset(cntr, 0, sizeof(cntr));
- for (n = 0; n < 17; n++) {
- res = get_bits(gb, 3);
- if (cntr[res] > 3) // should be >= 3 + (res == 7))
- return -1;
- vbm_tree[res * 3 + cntr[res]++] = n;
- }
INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),
bits, 1, 1, codes, 2, 2, 132);
- return 0;
}
/**
av_log(ctx, AV_LOG_ERROR,
"Invalid extradata size %d (should be 46)\n",
ctx->extradata_size);
- return -1;
+ return AVERROR_INVALIDDATA;
}
flags = AV_RL32(ctx->extradata + 18);
s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
s->do_apf = flags & 0x1;
+ if (s->do_apf) {
+ ff_rdft_init(&s->rdft, 7, DFT_R2C);
+ ff_rdft_init(&s->irdft, 7, IDFT_C2R);
+ ff_dct_init(&s->dct, 6, DCT_I);
+ ff_dct_init(&s->dst, 6, DST_I);
+
+ ff_sine_window_init(s->cos, 256);
+ memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
+ for (n = 0; n < 255; n++) {
+ s->sin[n] = -s->sin[510 - n];
+ s->cos[510 - n] = s->cos[n];
+ }
+ }
+ s->denoise_strength = (flags >> 2) & 0xF;
+ if (s->denoise_strength >= 12) {
+ av_log(ctx, AV_LOG_ERROR,
+ "Invalid denoise filter strength %d (max=11)\n",
+ s->denoise_strength);
+ return AVERROR_INVALIDDATA;
+ }
+ s->denoise_tilt_corr = !!(flags & 0x40);
+ s->dc_level = (flags >> 7) & 0xF;
s->lsp_q_mode = !!(flags & 0x2000);
s->lsp_def_mode = !!(flags & 0x4000);
lsp16_flag = flags & 0x1000;
init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
- return -1;
+ return AVERROR_INVALIDDATA;
}
s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
pitch_range = s->max_pitch_val - s->min_pitch_val;
+ if (pitch_range <= 0) {
+ av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n");
+ return AVERROR_INVALIDDATA;
+ }
s->pitch_nbits = av_ceil_log2(pitch_range);
s->last_pitch_val = 40;
s->last_acb_type = ACB_TYPE_NONE;
"Unsupported samplerate %d (min=%d, max=%d)\n",
ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
- return -1;
+ return AVERROR(ENOSYS);
}
s->block_conv_table[0] = s->min_pitch_val;
s->block_conv_table[2] = (pitch_range * 44) >> 6;
s->block_conv_table[3] = s->max_pitch_val - 1;
s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
+ if (s->block_delta_pitch_hrange <= 0) {
+ av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n");
+ return AVERROR_INVALIDDATA;
+ }
s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
s->block_pitch_range = s->block_conv_table[2] +
s->block_conv_table[3] + 1 +
2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
- ctx->sample_fmt = SAMPLE_FMT_FLT;
+ ctx->channels = 1;
+ ctx->channel_layout = AV_CH_LAYOUT_MONO;
+ ctx->sample_fmt = AV_SAMPLE_FMT_FLT;
+
+ return 0;
+}
+
+/**
+ * @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 = avpriv_scalarproduct_float_c(in, ptr, size);
+ if (dot > optimal_gain) {
+ optimal_gain = dot;
+ best_hist_ptr = ptr;
+ }
+ } while (--ptr >= end);
+
+ if (optimal_gain <= 0)
+ return -1;
+ dot = avpriv_scalarproduct_float_c(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 + avpriv_scalarproduct_float_c(lpcs, lpcs, n_lpcs);
+ rh1 = lpcs[0] + avpriv_scalarproduct_float_c(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) { // fall back 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 sine 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 / avpriv_scalarproduct_float_c(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);
+ }
+}
+/**
+ * @}
+ */
+
/**
* Dequantize LSPs
* @param lsps output pointer to the array that will hold the LSPs
}
/**
- * @defgroup lsp_dequant LSP dequantization routines
+ * @name LSP dequantization routines
* LSP dequantization routines, for 10/16LSPs and independent/residual coding.
* @note we assume enough bits are available, caller should check.
* lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
/**
* @}
- * @defgroup aw Pitch-adaptive window coding functions
+ * @name Pitch-adaptive window coding functions
* The next few functions are for pitch-adaptive window coding.
* @{
*/
* @param gb bit I/O context
* @param block_idx block index in frame [0, 1]
* @param fcb structure containing fixed codebook vector info
+ * @return -1 on error, 0 otherwise
*/
-static void aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
- int block_idx, AMRFixed *fcb)
+static int aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
+ int block_idx, AMRFixed *fcb)
{
- uint16_t use_mask[7]; // only 5 are used, rest is padding
+ uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
+ uint16_t *use_mask = use_mask_mem + 2;
/* in this function, idx is the index in the 80-bit (+ padding) use_mask
* bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
* of idx are the position of the bit within a particular item in the
/* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
* in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
* we exclude that range from being pulsed again in this function. */
+ memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
memset( use_mask, -1, 5 * sizeof(use_mask[0]));
memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
if (s->aw_n_pulses[block_idx] > 0)
int excl_range = s->aw_pulse_range; // always 16 or 24
uint16_t *use_mask_ptr = &use_mask[idx >> 4];
int first_sh = 16 - (idx & 15);
- *use_mask_ptr++ &= 0xFFFF << first_sh;
+ *use_mask_ptr++ &= 0xFFFFu << first_sh;
excl_range -= first_sh;
if (excl_range >= 16) {
*use_mask_ptr++ = 0;
else if (use_mask[2]) idx = 0x2F;
else if (use_mask[3]) idx = 0x3F;
else if (use_mask[4]) idx = 0x4F;
- else return;
+ else return -1;
idx -= av_log2_16bit(use_mask[idx >> 4]);
}
if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
/* set offset for next block, relative to start of that block */
n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
+ return 0;
}
/**
* (fixed) codebook pulses of the speech signal. */
if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
aw_pulse_set1(s, gb, block_idx, &fcb);
- aw_pulse_set2(s, gb, block_idx, &fcb);
+ if (aw_pulse_set2(s, gb, block_idx, &fcb)) {
+ /* Conceal the block with silence and return.
+ * Skip the correct amount of bits to read the next
+ * block from the correct offset. */
+ int r_idx = pRNG(s->frame_cntr, block_idx, size);
+
+ for (n = 0; n < size; n++)
+ excitation[n] =
+ wmavoice_std_codebook[r_idx + n] * s->silence_gain;
+ skip_bits(gb, 7 + 1);
+ return;
+ }
} else /* FCB_TYPE_EXC_PULSES */ {
int offset_nbits = 5 - frame_desc->log_n_blocks;
/* Calculate gain for adaptive & fixed codebook signal.
* see ff_amr_set_fixed_gain(). */
idx = get_bits(gb, 7);
- fcb_gain = expf(ff_dot_productf(s->gain_pred_err, gain_coeff, 6) -
+ fcb_gain = expf(avpriv_scalarproduct_float_c(s->gain_pred_err,
+ gain_coeff, 6) -
5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
acb_gain = wmavoice_gain_codebook_acb[idx];
pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
wmavoice_ipol2_coeffs, 4,
idx, 8, size);
} else
- av_memcpy_backptr(excitation, sizeof(float) * block_pitch,
+ av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
sizeof(float) * size);
}
*
* @param ctx WMA Voice decoder context
* @param gb bit I/O context (s->gb or one for cross-packet superframes)
+ * @param frame_idx Frame number within superframe [0-2]
* @param samples pointer to output sample buffer, has space for at least 160
* samples
* @param lsps LSP array
* @param synth target buffer for synthesized speech data
* @return 0 on success, <0 on error.
*/
-static int synth_frame(AVCodecContext *ctx, GetBitContext *gb,
+static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx,
float *samples,
const double *lsps, const double *prev_lsps,
float *excitation, float *synth)
int pitch[MAX_BLOCKS], last_block_pitch;
/* Parse frame type ("frame header"), see frame_descs */
- int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)],
- block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
+ int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)], block_nsamples;
if (bd_idx < 0) {
av_log(ctx, AV_LOG_ERROR,
"Invalid frame type VLC code, skipping\n");
- return -1;
+ return AVERROR_INVALIDDATA;
}
+ block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
+
/* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
/* Pitch is provided per frame, which is interpreted as the pitch of
/* Averaging projection filter, if applicable. Else, just copy samples
* from synthesis buffer */
if (s->do_apf) {
- // FIXME this is where APF would take place, currently not implemented
- av_log_missing_feature(ctx, "APF", 0);
- s->do_apf = 0;
- } //else
+ double i_lsps[MAX_LSPS];
+ float lpcs[MAX_LSPS];
+
+ for (n = 0; n < s->lsps; n++) // LSF -> LSP
+ i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
+ ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
+ postfilter(s, synth, samples, 80, lpcs,
+ &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
+ frame_descs[bd_idx].fcb_type, pitch[0]);
+
+ for (n = 0; n < s->lsps; n++) // LSF -> LSP
+ i_lsps[n] = cos(lsps[n]);
+ ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
+ postfilter(s, &synth[80], &samples[80], 80, lpcs,
+ &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
+ frame_descs[bd_idx].fcb_type, pitch[0]);
+ } else
memcpy(samples, synth, 160 * sizeof(synth[0]));
/* Cache values for next frame */
* does not modify the state of the bitreader; it
* only uses it to copy the current stream position
* @param s WMA Voice decoding context private data
- * @return -1 if unsupported, 1 on not enough bits or 0 if OK.
+ * @return < 0 on error, 1 on not enough bits or 0 if OK.
*/
static int check_bits_for_superframe(GetBitContext *orig_gb,
WMAVoiceContext *s)
if (get_bits_left(gb) < 14)
return 1;
if (!get_bits1(gb))
- return -1; // WMAPro-in-WMAVoice superframe
+ return AVERROR(ENOSYS); // WMAPro-in-WMAVoice superframe
if (get_bits1(gb)) skip_bits(gb, 12); // number of samples in superframe
if (s->has_residual_lsps) { // residual LSPs (for all frames)
if (get_bits_left(gb) < s->sframe_lsp_bitsize)
}
bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)];
if (bd_idx < 0)
- return -1; // invalid frame type VLC code
+ return AVERROR_INVALIDDATA; // invalid frame type VLC code
frame_desc = &frame_descs[bd_idx];
if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
if (get_bits_left(gb) < s->pitch_nbits)
* (if less than 480), usually used to prevent blanks at track boundaries.
*
* @param ctx WMA Voice decoder context
- * @param samples pointer to output buffer for voice samples
- * @param data_size pointer containing the size of #samples on input, and the
- * amount of #samples filled on output
* @return 0 on success, <0 on error or 1 if there was not enough data to
* fully parse the superframe
*/
-static int synth_superframe(AVCodecContext *ctx,
- float *samples, int *data_size)
+static int synth_superframe(AVCodecContext *ctx, AVFrame *frame,
+ int *got_frame_ptr)
{
WMAVoiceContext *s = ctx->priv_data;
GetBitContext *gb = &s->gb, s_gb;
wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
float synth[MAX_LSPS + MAX_SFRAMESIZE];
+ float *samples;
memcpy(synth, s->synth_history,
s->lsps * sizeof(*synth));
s->sframe_cache_size = 0;
}
- if ((res = check_bits_for_superframe(gb, s)) == 1) return 1;
+ if ((res = check_bits_for_superframe(gb, s)) == 1) {
+ *got_frame_ptr = 0;
+ return 1;
+ } else if (res < 0)
+ return res;
/* First bit is speech/music bit, it differentiates between WMAVoice
* speech samples (the actual codec) and WMAVoice music samples, which
* are really WMAPro-in-WMAVoice-superframes. I've never seen those in
* the wild yet. */
if (!get_bits1(gb)) {
- av_log_missing_feature(ctx, "WMAPro-in-WMAVoice support", 1);
- return -1;
+ avpriv_request_sample(ctx, "WMAPro-in-WMAVoice");
+ return AVERROR_PATCHWELCOME;
}
/* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
av_log(ctx, AV_LOG_ERROR,
"Superframe encodes >480 samples (%d), not allowed\n",
n_samples);
- return -1;
+ return AVERROR_INVALIDDATA;
}
}
/* Parse LSPs, if global for the superframe (can also be per-frame). */
stabilize_lsps(lsps[n], s->lsps);
}
- /* Parse frames, optionally preceeded by per-frame (independent) LSPs. */
+ /* get output buffer */
+ frame->nb_samples = 480;
+ if ((res = ff_get_buffer(ctx, frame, 0)) < 0) {
+ av_log(ctx, AV_LOG_ERROR, "get_buffer() failed\n");
+ return res;
+ }
+ frame->nb_samples = n_samples;
+ samples = (float *)frame->data[0];
+
+ /* Parse frames, optionally preceded by per-frame (independent) LSPs. */
for (n = 0; n < 3; n++) {
if (!s->has_residual_lsps) {
int m;
stabilize_lsps(lsps[n], s->lsps);
}
- if ((res = synth_frame(ctx, gb,
+ if ((res = synth_frame(ctx, gb, n,
&samples[n * MAX_FRAMESIZE],
lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
&excitation[s->history_nsamples + n * MAX_FRAMESIZE],
- &synth[s->lsps + n * MAX_FRAMESIZE])))
+ &synth[s->lsps + n * MAX_FRAMESIZE]))) {
+ *got_frame_ptr = 0;
return res;
+ }
}
/* Statistics? FIXME - we don't check for length, a slight overrun
skip_bits(gb, 10 * (res + 1));
}
- /* Specify nr. of output samples */
- *data_size = n_samples * sizeof(float);
+ *got_frame_ptr = 1;
/* Update history */
memcpy(s->prev_lsps, lsps[2],
s->lsps * sizeof(*synth));
memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
s->history_nsamples * sizeof(*excitation));
+ if (s->do_apf)
+ memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
+ s->history_nsamples * sizeof(*s->zero_exc_pf));
return 0;
}
* @param size size of the source data, in bytes
* @param gb bit I/O context specifying the current position in the source.
* data. This function might use this to align the bit position to
- * a whole-byte boundary before calling #ff_copy_bits() on aligned
+ * a whole-byte boundary before calling #avpriv_copy_bits() on aligned
* source data
* @param nbits the amount of bits to copy from source to target
*
rmn_bits = rmn_bytes = get_bits_left(gb);
if (rmn_bits < nbits)
return;
+ if (nbits > pb->size_in_bits - put_bits_count(pb))
+ return;
rmn_bits &= 7; rmn_bytes >>= 3;
if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
- ff_copy_bits(pb, data + size - rmn_bytes,
+ avpriv_copy_bits(pb, data + size - rmn_bytes,
FFMIN(nbits - rmn_bits, rmn_bytes << 3));
}
* For more information about frames, see #synth_superframe().
*/
static int wmavoice_decode_packet(AVCodecContext *ctx, void *data,
- int *data_size, AVPacket *avpkt)
+ int *got_frame_ptr, AVPacket *avpkt)
{
WMAVoiceContext *s = ctx->priv_data;
GetBitContext *gb = &s->gb;
int size, res, pos;
- if (*data_size < 480 * sizeof(float)) {
- av_log(ctx, AV_LOG_ERROR,
- "Output buffer too small (%d given - %lu needed)\n",
- *data_size, 480 * sizeof(float));
- return -1;
- }
- *data_size = 0;
-
/* Packets are sometimes a multiple of ctx->block_align, with a packet
- * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer
+ * header at each ctx->block_align bytes. However, Libav's ASF demuxer
* feeds us ASF packets, which may concatenate multiple "codec" packets
* in a single "muxer" packet, so we artificially emulate that by
* capping the packet size at ctx->block_align. */
for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
- if (!size)
+ if (!size) {
+ *got_frame_ptr = 0;
return 0;
+ }
init_get_bits(&s->gb, avpkt->data, size << 3);
/* size == ctx->block_align is used to indicate whether we are dealing with
copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits);
flush_put_bits(&s->pb);
s->sframe_cache_size += s->spillover_nbits;
- if ((res = synth_superframe(ctx, data, data_size)) == 0 &&
- *data_size > 0) {
+ if ((res = synth_superframe(ctx, data, got_frame_ptr)) == 0 &&
+ *got_frame_ptr) {
cnt += s->spillover_nbits;
s->skip_bits_next = cnt & 7;
return cnt >> 3;
s->sframe_cache_size = 0;
s->skip_bits_next = 0;
pos = get_bits_left(gb);
- if ((res = synth_superframe(ctx, data, data_size)) < 0) {
+ if ((res = synth_superframe(ctx, data, got_frame_ptr)) < 0) {
return res;
- } else if (*data_size > 0) {
+ } else if (*got_frame_ptr) {
int cnt = get_bits_count(gb);
s->skip_bits_next = cnt & 7;
return cnt >> 3;
return size;
}
+static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
+{
+ WMAVoiceContext *s = ctx->priv_data;
+
+ if (s->do_apf) {
+ ff_rdft_end(&s->rdft);
+ ff_rdft_end(&s->irdft);
+ ff_dct_end(&s->dct);
+ ff_dct_end(&s->dst);
+ }
+
+ return 0;
+}
+
static av_cold void wmavoice_flush(AVCodecContext *ctx)
{
WMAVoiceContext *s = ctx->priv_data;
int n;
+ s->postfilter_agc = 0;
s->sframe_cache_size = 0;
s->skip_bits_next = 0;
for (n = 0; n < s->lsps; n++)
sizeof(*s->synth_history) * MAX_LSPS);
memset(s->gain_pred_err, 0,
sizeof(s->gain_pred_err));
+
+ if (s->do_apf) {
+ memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
+ sizeof(*s->synth_filter_out_buf) * s->lsps);
+ memset(s->dcf_mem, 0,
+ sizeof(*s->dcf_mem) * 2);
+ memset(s->zero_exc_pf, 0,
+ sizeof(*s->zero_exc_pf) * s->history_nsamples);
+ memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
+ }
}
-AVCodec wmavoice_decoder = {
- "wmavoice",
- AVMEDIA_TYPE_AUDIO,
- CODEC_ID_WMAVOICE,
- sizeof(WMAVoiceContext),
- wmavoice_decode_init,
- NULL,
- NULL,
- wmavoice_decode_packet,
- CODEC_CAP_SUBFRAMES,
- .flush = wmavoice_flush,
- .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),
+AVCodec ff_wmavoice_decoder = {
+ .name = "wmavoice",
+ .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),
+ .type = AVMEDIA_TYPE_AUDIO,
+ .id = AV_CODEC_ID_WMAVOICE,
+ .priv_data_size = sizeof(WMAVoiceContext),
+ .init = wmavoice_decode_init,
+ .init_static_data = wmavoice_init_static_data,
+ .close = wmavoice_decode_end,
+ .decode = wmavoice_decode_packet,
+ .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
+ .flush = wmavoice_flush,
};
projects / ffmpeg / blobdiff