「パケット損失隠蔽」を使用して、オーディオ ストリーム内の失われた PCM フレームを隠蔽しようとしています。残念ながら、すべてのライセンス制限とコードの肥大化なしにアクセスできるライブラリを見つけることができません (...いくつかの提案があります)。
PLC を実装する Asterisk プロジェクトのために Steve Underwood によって書かれたいくつかの GPL コードを見つけました。いくつかの制限があります。ただし、Steve が彼のコードで示唆しているように、彼のアルゴリズムは、少し手を加えるだけでさまざまなストリームに適用できます。現在、コードは 8kHz 16 ビット符号付きモノラル ストリームで動作します。
コードのバリエーションは、 Google Code Searchの簡単な検索で見つけることができます。
私の希望は、コードを他のストリームで動作するように適応させることです。最初の目標は、アルゴリズムを 8+ kHz、16 ビット符号付き、マルチチャンネル オーディオ (すべて C++ 環境) に合わせることです。最終的には、GPL ライセンスの下でコードを利用できるようにすることを検討しており、それが他の人の利益になることを期待しています...
私の努力による以下のコードが添付されています。コードには、特定の確率で多数のフレームを「ドロップ」するメイン関数が含まれています。残念ながら、コードは期待どおりには機能しません。gdb で実行すると EXC_BAD_ACCESS を受け取りますが、「bt」コマンドを使用すると gdb からトレースが取得されません。明らかに、どこかでメモリを踏みにじっていますが、正確な場所はわかりません。amdf_pitch関数をコメントアウトすると、コードはクラッシュせずに実行されます...
int main (int argc, char *argv[])
{
std::ifstream fin("C:\\cc32kHz.pcm");
if(!fin.is_open())
{
std::cout << "Failed to open input file" << std::endl;
return 1;
}
std::ofstream fout_repaired("C:\\cc32kHz_repaired.pcm");
if(!fout_repaired.is_open())
{
std::cout << "Failed to open output repaired file" << std::endl;
return 1;
}
std::ofstream fout_lossy("C:\\cc32kHz_lossy.pcm");
if(!fout_lossy.is_open())
{
std::cout << "Failed to open output repaired file" << std::endl;
return 1;
}
audio::PcmConcealer Concealer;
Concealer.Init(1, 16, 32000);
//Generate random numbers;
srand( time(NULL) );
int value = 0;
int probability = 5;
while(!fin.eof())
{
char arr[2];
fin.read(arr, 2);
//Generate's random number;
value = rand() % 100 + 1;
if(value <= probability)
{
char blank[2] = {0x00, 0x00};
fout_lossy.write(blank, 2);
//Fill in data;
Concealer.Fill((int16_t *)blank, 1);
fout_repaired.write(blank, 2);
}
else
{
//Write data to file;
fout_repaired.write(arr, 2);
fout_lossy.write(arr, 2);
Concealer.Receive((int16_t *)arr, 1);
}
}
fin.close();
fout_repaired.close();
fout_lossy.close();
return 0;
}
PcmConcealer.hpp
/*
* Code adapted from Steve Underwood of the Asterisk Project. This code inherits
* the same licensing restrictions as the Asterisk Project.
*/
#ifndef __PCMCONCEALER_HPP__
#define __PCMCONCEALER_HPP__
/**
1. What does it do?
The packet loss concealment module provides a suitable synthetic fill-in signal,
to minimise the audible effect of lost packets in VoIP applications. It is not
tied to any particular codec, and could be used with almost any codec which does not
specify its own procedure for packet loss concealment.
Where a codec specific concealment procedure exists, the algorithm is usually built
around knowledge of the characteristics of the particular codec. It will, therefore,
generally give better results for that particular codec than this generic concealer will.
2. How does it work?
While good packets are being received, the plc_rx() routine keeps a record of the trailing
section of the known speech signal. If a packet is missed, plc_fillin() is called to produce
a synthetic replacement for the real speech signal. The average mean difference function
(AMDF) is applied to the last known good signal, to determine its effective pitch.
Based on this, the last pitch period of signal is saved. Essentially, this cycle of speech
will be repeated over and over until the real speech resumes. However, several refinements
are needed to obtain smooth pleasant sounding results.
- The two ends of the stored cycle of speech will not always fit together smoothly. This can
cause roughness, or even clicks, at the joins between cycles. To soften this, the
1/4 pitch period of real speech preceeding the cycle to be repeated is blended with the last
1/4 pitch period of the cycle to be repeated, using an overlap-add (OLA) technique (i.e.
in total, the last 5/4 pitch periods of real speech are used).
- The start of the synthetic speech will not always fit together smoothly with the tail of
real speech passed on before the erasure was identified. Ideally, we would like to modify
the last 1/4 pitch period of the real speech, to blend it into the synthetic speech. However,
it is too late for that. We could have delayed the real speech a little, but that would
require more buffer manipulation, and hurt the efficiency of the no-lost-packets case
(which we hope is the dominant case). Instead we use a degenerate form of OLA to modify
the start of the synthetic data. The last 1/4 pitch period of real speech is time reversed,
and OLA is used to blend it with the first 1/4 pitch period of synthetic speech. The result
seems quite acceptable.
- As we progress into the erasure, the chances of the synthetic signal being anything like
correct steadily fall. Therefore, the volume of the synthesized signal is made to decay
linearly, such that after 50ms of missing audio it is reduced to silence.
- When real speech resumes, an extra 1/4 pitch period of sythetic speech is blended with the
start of the real speech. If the erasure is small, this smoothes the transition. If the erasure
is long, and the synthetic signal has faded to zero, the blending softens the start up of the
real signal, avoiding a kind of "click" or "pop" effect that might occur with a sudden onset.
3. How do I use it?
Before audio is processed, call plc_init() to create an instance of the packet loss
concealer. For each received audio packet that is acceptable (i.e. not including those being
dropped for being too late) call plc_rx() to record the content of the packet. Note this may
modify the packet a little after a period of packet loss, to blend real synthetic data smoothly.
When a real packet is not available in time, call plc_fillin() to create a sythetic substitute.
That's it!
*/
/*! Minimum allowed pitch (66 Hz) */
#define PLC_PITCH_MIN(SAMPLE_RATE) ((double)(SAMPLE_RATE) / 66.6)
/*! Maximum allowed pitch (200 Hz) */
#define PLC_PITCH_MAX(SAMPLE_RATE) ((SAMPLE_RATE) / 200)
/*! Maximum pitch OLA window */
//#define PLC_PITCH_OVERLAP_MAX(SAMPLE_RATE) ((PLC_PITCH_MIN(SAMPLE_RATE)) >> 2)
/*! The length over which the AMDF function looks for similarity (20 ms) */
#define CORRELATION_SPAN(SAMPLE_RATE) ((20 * (SAMPLE_RATE)) / 1000)
/*! History buffer length. The buffer must also be at leat 1.25 times
PLC_PITCH_MIN, but that is much smaller than the buffer needs to be for
the pitch assessment. */
//#define PLC_HISTORY_LEN(SAMPLE_RATE) ((CORRELATION_SPAN(SAMPLE_RATE)) + (PLC_PITCH_MIN(SAMPLE_RATE)))
namespace audio
{
typedef struct
{
/*! Consecutive erased samples */
int missing_samples;
/*! Current offset into pitch period */
int pitch_offset;
/*! Pitch estimate */
int pitch;
/*! Buffer for a cycle of speech */
float *pitchbuf;//[PLC_PITCH_MIN];
/*! History buffer */
short *history;//[PLC_HISTORY_LEN];
/*! Current pointer into the history buffer */
int buf_ptr;
} plc_state_t;
class PcmConcealer
{
public:
PcmConcealer();
~PcmConcealer();
void Init(int channels, int bit_depth, int sample_rate);
//Process a block of received audio samples.
int Receive(short amp[], int frames);
//Fill-in a block of missing audio samples.
int Fill(short amp[], int frames);
void Destroy();
private:
int amdf_pitch(int min_pitch, int max_pitch, short amp[], int channel_index, int frames);
void save_history(plc_state_t *s, short *buf, int channel_index, int frames);
void normalise_history(plc_state_t *s);
/** Holds the states of each of the channels **/
std::vector< plc_state_t * > ChannelStates;
int plc_pitch_min;
int plc_pitch_max;
int plc_pitch_overlap_max;
int correlation_span;
int plc_history_len;
int channel_count;
int sample_rate;
bool Initialized;
};
}
#endif
Pcmコンシーラー.cpp
/*
* Code adapted from Steve Underwood of the Asterisk Project. This code inherits
* the same licensing restrictions as the Asterisk Project.
*/
#include "audio/PcmConcealer.hpp"
/* We do a straight line fade to zero volume in 50ms when we are filling in for missing data. */
#define ATTENUATION_INCREMENT 0.0025 /* Attenuation per sample */
#if !defined(INT16_MAX)
#define INT16_MAX (32767)
#define INT16_MIN (-32767-1)
#endif
#ifdef WIN32
inline double rint(double x)
{
return floor(x + 0.5);
}
#endif
inline short fsaturate(double damp)
{
if (damp > 32767.0)
return INT16_MAX;
if (damp < -32768.0)
return INT16_MIN;
return (short)rint(damp);
}
namespace audio
{
PcmConcealer::PcmConcealer() : Initialized(false)
{
}
PcmConcealer::~PcmConcealer()
{
Destroy();
}
void PcmConcealer::Init(int channels, int bit_depth, int sample_rate)
{
if(Initialized)
return;
if(channels <= 0 || bit_depth != 16)
return;
Initialized = true;
channel_count = channels;
this->sample_rate = sample_rate;
//////////////
double min = PLC_PITCH_MIN(sample_rate);
int imin = (int)min;
double max = PLC_PITCH_MAX(sample_rate);
int imax = (int)max;
plc_pitch_min = imin;
plc_pitch_max = imax;
plc_pitch_overlap_max = (plc_pitch_min >> 2);
correlation_span = CORRELATION_SPAN(sample_rate);
plc_history_len = correlation_span + plc_pitch_min;
//////////////
for(int i = 0; i < channel_count; i ++)
{
plc_state_t *t = new plc_state_t;
memset(t, 0, sizeof(plc_state_t));
t->pitchbuf = new float[plc_pitch_min];
t->history = new short[plc_history_len];
ChannelStates.push_back(t);
}
}
void PcmConcealer::Destroy()
{
if(!Initialized)
return;
while(ChannelStates.size())
{
plc_state_t *s = ChannelStates.at(0);
if(s)
{
if(s->history) delete s->history;
if(s->pitchbuf) delete s->pitchbuf;
memset(s, 0, sizeof(plc_state_t));
delete s;
}
ChannelStates.erase(ChannelStates.begin());
}
ChannelStates.clear();
Initialized = false;
}
//Process a block of received audio samples.
int PcmConcealer::Receive(short amp[], int frames)
{
if(!Initialized)
return 0;
int j = 0;
for(int k = 0; k < ChannelStates.size(); k++)
{
int i;
int overlap_len;
int pitch_overlap;
float old_step;
float new_step;
float old_weight;
float new_weight;
float gain;
plc_state_t *s = ChannelStates.at(k);
if (s->missing_samples)
{
/* Although we have a real signal, we need to smooth it to fit well
with the synthetic signal we used for the previous block */
/* The start of the real data is overlapped with the next 1/4 cycle
of the synthetic data. */
pitch_overlap = s->pitch >> 2;
if (pitch_overlap > frames)
pitch_overlap = frames;
gain = 1.0 - s->missing_samples * ATTENUATION_INCREMENT;
if (gain < 0.0)
gain = 0.0;
new_step = 1.0/pitch_overlap;
old_step = new_step*gain;
new_weight = new_step;
old_weight = (1.0 - new_step)*gain;
for (i = 0; i < pitch_overlap; i++)
{
int index = (i * channel_count) + j;
amp[index] = fsaturate(old_weight * s->pitchbuf[s->pitch_offset] + new_weight * amp[index]);
if (++s->pitch_offset >= s->pitch)
s->pitch_offset = 0;
new_weight += new_step;
old_weight -= old_step;
if (old_weight < 0.0)
old_weight = 0.0;
}
s->missing_samples = 0;
}
save_history(s, amp, j, frames);
j++;
}
return frames;
}
//Fill-in a block of missing audio samples.
int PcmConcealer::Fill(short amp[], int frames)
{
if(!Initialized)
return 0;
int j =0;
for(int k = 0; k < ChannelStates.size(); k++)
{
short *tmp = new short[plc_pitch_overlap_max];
int i;
int pitch_overlap;
float old_step;
float new_step;
float old_weight;
float new_weight;
float gain;
short *orig_amp;
int orig_len;
orig_amp = amp;
orig_len = frames;
plc_state_t *s = ChannelStates.at(k);
if (s->missing_samples == 0)
{
// As the gap in real speech starts we need to assess the last known pitch,
//and prepare the synthetic data we will use for fill-in
normalise_history(s);
s->pitch = amdf_pitch(plc_pitch_min, plc_pitch_max, s->history + plc_history_len - correlation_span - plc_pitch_min, j, correlation_span);
// We overlap a 1/4 wavelength
pitch_overlap = s->pitch >> 2;
// Cook up a single cycle of pitch, using a single of the real signal with 1/4
//cycle OLA'ed to make the ends join up nicely
// The first 3/4 of the cycle is a simple copy
for (i = 0; i < s->pitch - pitch_overlap; i++)
s->pitchbuf[i] = s->history[plc_history_len - s->pitch + i];
// The last 1/4 of the cycle is overlapped with the end of the previous cycle
new_step = 1.0/pitch_overlap;
new_weight = new_step;
for ( ; i < s->pitch; i++)
{
s->pitchbuf[i] = s->history[plc_history_len - s->pitch + i]*(1.0 - new_weight) + s->history[plc_history_len - 2*s->pitch + i]*new_weight;
new_weight += new_step;
}
// We should now be ready to fill in the gap with repeated, decaying cycles
// of what is in pitchbuf
// We need to OLA the first 1/4 wavelength of the synthetic data, to smooth
// it into the previous real data. To avoid the need to introduce a delay
// in the stream, reverse the last 1/4 wavelength, and OLA with that.
gain = 1.0;
new_step = 1.0/pitch_overlap;
old_step = new_step;
new_weight = new_step;
old_weight = 1.0 - new_step;
for (i = 0; i < pitch_overlap; i++)
{
int index = (i * channel_count) + j;
amp[index] = fsaturate(old_weight * s->history[plc_history_len - 1 - i] + new_weight * s->pitchbuf[i]);
new_weight += new_step;
old_weight -= old_step;
if (old_weight < 0.0)
old_weight = 0.0;
}
s->pitch_offset = i;
}
else
{
gain = 1.0 - s->missing_samples*ATTENUATION_INCREMENT;
i = 0;
}
for ( ; gain > 0.0 && i < frames; i++)
{
int index = (i * channel_count) + j;
amp[index] = s->pitchbuf[s->pitch_offset]*gain;
gain -= ATTENUATION_INCREMENT;
if (++s->pitch_offset >= s->pitch)
s->pitch_offset = 0;
}
for ( ; i < frames; i++)
{
int index = (i * channel_count) + j;
amp[i] = 0;
}
s->missing_samples += orig_len;
save_history(s, amp, j, frames);
delete [] tmp;
j++;
}
return frames;
}
void PcmConcealer::save_history(plc_state_t *s, short *buf, int channel_index, int frames)
{
if (frames >= plc_history_len)
{
/* Just keep the last part of the new data, starting at the beginning of the buffer */
//memcpy(s->history, buf + len - plc_history_len, sizeof(short)*plc_history_len);
int frames_to_copy = plc_history_len;
for(int i = 0; i < frames_to_copy; i ++)
{
int index = (channel_count * (i + frames - plc_history_len)) + channel_index;
s->history[i] = buf[index];
}
s->buf_ptr = 0;
return;
}
if (s->buf_ptr + frames > plc_history_len)
{
/* Wraps around - must break into two sections */
//memcpy(s->history + s->buf_ptr, buf, sizeof(short)*(plc_history_len - s->buf_ptr));
short *hist_ptr = s->history + s->buf_ptr;
int frames_to_copy = plc_history_len - s->buf_ptr;
for(int i = 0; i < frames_to_copy; i ++)
{
int index = (channel_count * i) + channel_index;
hist_ptr[i] = buf[index];
}
frames -= (plc_history_len - s->buf_ptr);
//memcpy(s->history, buf + (plc_history_len - s->buf_ptr), sizeof(short)*len);
frames_to_copy = frames;
for(int i = 0; i < frames_to_copy; i ++)
{
int index = (channel_count * (i + (plc_history_len - s->buf_ptr))) + channel_index;
s->history[i] = buf[index];
}
s->buf_ptr = frames;
return;
}
/* Can use just one section */
//memcpy(s->history + s->buf_ptr, buf, sizeof(short)*len);
short *hist_ptr = s->history + s->buf_ptr;
int frames_to_copy = frames;
for(int i = 0; i < frames_to_copy; i ++)
{
int index = (channel_count * i) + channel_index;
hist_ptr[i] = buf[index];
}
s->buf_ptr += frames;
}
void PcmConcealer::normalise_history(plc_state_t *s)
{
short *tmp = new short[plc_history_len];
if (s->buf_ptr == 0)
return;
memcpy(tmp, s->history, sizeof(short)*s->buf_ptr);
memcpy(s->history, s->history + s->buf_ptr, sizeof(short)*(plc_history_len - s->buf_ptr));
memcpy(s->history + plc_history_len - s->buf_ptr, tmp, sizeof(short)*s->buf_ptr);
s->buf_ptr = 0;
delete [] tmp;
}
int PcmConcealer::amdf_pitch(int min_pitch, int max_pitch, short amp[], int channel_index, int frames)
{
int i;
int j;
int acc;
int min_acc;
int pitch;
pitch = min_pitch;
min_acc = INT_MAX;
for (i = max_pitch; i <= min_pitch; i++)
{
acc = 0;
for (j = 0; j < frames; j++)
{
int index1 = (channel_count * (i+j)) + channel_index;
int index2 = (channel_count * j) + channel_index;
//std::cout << "Index 1: " << index1 << ", Index 2: " << index2 << std::endl;
acc += abs(amp[index1] - amp[index2]);
}
if (acc < min_acc)
{
min_acc = acc;
pitch = i;
}
}
std::cout << "Pitch: " << pitch << std::endl;
return pitch;
}
}
PS - デジタル オーディオは私の得意分野ではないことを告白しなければなりません...