Radio ad blocker
0 likes157 views
This document describes the development of a radio ad blocker that uses digital signal processing to detect and mute broadcast commercials in real-time. It discusses analyzing the audio stream to find matching segments, reframes the problem to compare short sound samples, prototypes approaches using octave like cross-correlation on waveforms and FFTs, and describes a Java implementation that analyzes a sliding window to identify matching jingles marking ad breaks.
![Reading PCM data
byte[] rawBuffer = new byte[4096];
final int SAMPLE_SIZE = 4; // 2 bytes * 2 channels
ByteBuffer sample = ByteBuffer.allocate(SAMPLE_SIZE);
sample.order(ByteOrder.LITTLE_ENDIAN);
while (true) {
int length = input.read(rawBuffer);
// consume the rawBuffer
for (int i = 0; i < length; i += SAMPLE_SIZE) {
sample.clear();
sample.put(rawBuffer, i, SAMPLE_SIZE);
sample.rewind();
while (sample.hasRemaining()) {
short left = sample.getShort();
short right = sample.getShort();
// ...
}
}
}](https://image.slidesharecdn.com/radioadblocker-200323073149/85/Radio-ad-blocker-10-320.jpg)
![Porting xcorr from Octave
N = max(length(X),length(Y));
maxlag = N - 1;
M = 2^nextpow2(N + maxlag);
pre = fft( postpad( prepad( X(:), length(X)+maxlag ), M) );
post = fft( postpad( Y(:), M ) );
cor = ifft( pre .* conj(post) );
R = cor(1:2*maxlag+1);
R = real(R);
public class Waveform {
private final float[] buffer;
private final int length;
// ...
}
Waveform x = currentWindow;
Waveform y = jingle
.setPadding(jingle.getSize() + windowSize - 1, next2Pow)
.doFft();
x.doFft();
x.doConjAndMultiply(y);
x.doIfft();
float[] result = y.getMaxReal(2 * windowSize + 1);](https://image.slidesharecdn.com/radioadblocker-200323073149/85/Radio-ad-blocker-13-320.jpg)
![FFT usage in Java
// https://github.com/wendykierp/JTransforms
import org.jtransforms.fft.FloatFFT_1D;
short[] shortBuffer; // (-32678,32767)
float[] buffer = shortBuffer / 32768;
// fft()
FloatFFT_1D fft = new FloatFFT_1D(length);
fft.realForwardFull(buffer);
// ifft()
fft.complexInverse(buffer, true);
buffer[i]; // i%2 == 0; real part
buffer[i+1]; // imaginary part](https://image.slidesharecdn.com/radioadblocker-200323073149/85/Radio-ad-blocker-14-320.jpg)
Radio ad blocker
- 1. Radio ad blocker Or how I learned to stop worrying and use DSP cargo cult to mute the broadcast commercials
- 2. Analysing radio broadcast Starting jingle Ad 1 Ad 2 Ad 3 Ending jingle Volume controlSample analyser Broadcast
- 3. Research process: finding a sample in the sound stream
- 4. Redefining the problem • Let’s make it a simpler problem, by taking away the part that we can figure out ourselves • We can split the incoming stream into sound samples of some length (moving window) • So let’s try to find a way to compare two sound samples
- 7. Cross correlation Matching samples Not matching samples
- 8. Prototyping with Octave pkg load signal; stream = audioread('stream.wav'); jingle1 = audioread('commercial-start.wav'); jingle2 = audioread('commercial-end.wav'); function retval = get_sample(data, from, to) retval = data(from*32000 : to*32000-1); endfunction sample = get_sample(stream, 134, 135.5); plot(sample); plot(jingle2); plot(real(fft(sample))); plot(real(fft(jingle2))); plot(xcorr(sample, jingle2));
- 9. Java implementation Analyser Broadcast PCM stdout jingle 1 stdin stderrrtl_fm -M wbfm -f 89.50M -g 0 -r 48000 ffmpeg -i http://41.dktr.pl:8000/trojka.ogg -f s16le -acodec pcm_s16le - play -r 48000 -b 16 -c 2 -e signed -t raw -
- 10. Reading PCM data byte[] rawBuffer = new byte[4096]; final int SAMPLE_SIZE = 4; // 2 bytes * 2 channels ByteBuffer sample = ByteBuffer.allocate(SAMPLE_SIZE); sample.order(ByteOrder.LITTLE_ENDIAN); while (true) { int length = input.read(rawBuffer); // consume the rawBuffer for (int i = 0; i < length; i += SAMPLE_SIZE) { sample.clear(); sample.put(rawBuffer, i, SAMPLE_SIZE); sample.rewind(); while (sample.hasRemaining()) { short left = sample.getShort(); short right = sample.getShort(); // ... } } }
- 13. Porting xcorr from Octave N = max(length(X),length(Y)); maxlag = N - 1; M = 2^nextpow2(N + maxlag); pre = fft( postpad( prepad( X(:), length(X)+maxlag ), M) ); post = fft( postpad( Y(:), M ) ); cor = ifft( pre .* conj(post) ); R = cor(1:2*maxlag+1); R = real(R); public class Waveform { private final float[] buffer; private final int length; // ... } Waveform x = currentWindow; Waveform y = jingle .setPadding(jingle.getSize() + windowSize - 1, next2Pow) .doFft(); x.doFft(); x.doConjAndMultiply(y); x.doIfft(); float[] result = y.getMaxReal(2 * windowSize + 1);
- 14. FFT usage in Java // https://github.com/wendykierp/JTransforms import org.jtransforms.fft.FloatFFT_1D; short[] shortBuffer; // (-32678,32767) float[] buffer = shortBuffer / 32768; // fft() FloatFFT_1D fft = new FloatFFT_1D(length); fft.realForwardFull(buffer); // ifft() fft.complexInverse(buffer, true); buffer[i]; // i%2 == 0; real part buffer[i+1]; // imaginary part
- 15. Toolbox • sox package for uncompressed files (wav, raw, etc.) • sox - conversion • play - playback • ffmpeg - decoding compressed streams/files (ogg, mp3) • Octave • Audacity