Welcome, Guest

Want to take part in these discussions? Sign in if you have an account, or apply for one below

• Forgot your password?
• Apply for membership

1. • CommentAuthormlew
   • CommentTimeJun 14th 2010

    
   I'm working on a filtering application using microphone input through DirectSound. Whenever I run the ILMath.fft function on my array of data from the microphone, I get an array of junk data (NaN + NaNi). I've already tried replacing any NaN values in the input data with 0's, and that hasn't worked. Any ideas why this is happening?
2. • CommentAuthorhaymo
   • CommentTimeJun 15th 2010

    

   some ideas:

   1) try replacing prior NaNs with MachineParameterDouble.eps instead of 0

   2) find out, why any NaN exist at all (they shouldn't be there, I suppose?)

   3) make sure, you are using the proper datatype (single ? double? complex double? etc... )

   4) make sure, the array is properly detached from the DS buffer (by COPYING from the buffer or by cloning the array)

   If nothing helps, please post your code here and/or prepare a small test case for reference and reproducability. (I doubt, the reason is in the fft)

3. • CommentAuthormlew
   • CommentTimeJun 27th 2010

    
   Thanks for the tips - sorry it took so long to reply; I've been tied up with some other work. I tried replacing prior NaNs with MachineParameterDouble.eps, but that didn't work either. Not quite sure why those NaNs exist - all I'm doing is getting an array of data by reading the directsound buffer. Here's the code I've been working on (at least the relevant bits of it):
   
   private double[] processData(double[] data)
   {
   //create new ILArray from data to process
   ILArray<double> dataToProcess = new ILArray<double>(data);
   
   //compute 1D FFT on dataToProcess
   ILArray<ILNumerics.complex> processedData = ILMath.fft(dataToProcess);
   
   //get length of processedData
   int theLength = processedData.Length / 2;
   for (int i = 0; i < theLength; i++)
   {
   ILNumerics.complex c = (ILNumerics.complex)processedData[i];
   //get frequency at i
   double freq = Common.sampleRate * i / theLength;
   if (freq >= Common.lBound && freq <= Common.uBound) // amplify this region
   {
   c.real /= 2;
   }
   }
   ILArray<double> fData = ILMath.abs(ILMath.ifft(processedData));
   double[] final = fData.Values.ToArray<double>();
   
   return final; //change this later
   }
   
   private void EchoThread()
   {
   while (true)
   {
   Common.notificationEvent.WaitOne(Timeout.Infinite, true);
   
   byte[] theBuf = (byte[])buffer.Read(offset, typeof(byte), LockFlag.None, Common.notifySize);
   double[] dblBuf = (double[])buffer.Read(offset, typeof(double), LockFlag.None, Common.notifySize);
   for (int i = 0; i < dblBuf.Length; i++)
   {
   if (Double.IsNaN(dblBuf[i]))
   dblBuf[i] = new MachineParameterDouble().eps;
   }
   //process data
   double[] procBuf = processData(dblBuf);
   //System.Buffer.BlockCopy(procBuf, 0, theBuf, 0, theBuf.Length);
   byte[] fprocBuf = new byte[procBuf.Length];
   for (int i = 0; i < procBuf.Length; i++)
   {
   if (Double.IsNaN(procBuf[i]))
   fprocBuf[i] = (byte)0;
   else
   fprocBuf[i] = byte.Parse(procBuf[i].ToString());
   }
   //write contents of capture buffer to playback buffer
   Common.ampBuffer.Write(offset, fprocBuf, LockFlag.None);
   //move capture offset
   offset = (offset + Common.notifySize) % (Common.bufferPositions * Common.notifySize);
   //write captured sound to wav file and update length of wave file
   Writer.Write(theBuf, 0, theBuf.Length);
   length += theBuf.Length;
   }
   }
4. • CommentAuthorhaymo
   • CommentTimeJun 27th 2010

    

   Just some thoughts on possible optimizations... :

   private double[] processData(double[] data) //why not take an ILArray<double> here?
   {
   //create new ILArray from data to process
   ILArray<double> dataToProcess = new ILArray<double>(data);
   
   //compute 1D FFT on dataToProcess
   ILArray<ILNumerics.complex> processedData = ILMath.fft(dataToProcess);
   
   //get length of processedData

   // // the loop can be replaced by:
   // processedData[startIndex + ":" + endIndex] /= 2.0;
   // // (precomputed startIndex, endIndex)
   
   int theLength = processedData.Length / 2;
   for (int i = 0; i < theLength; i++)
   {
   ILNumerics.complex c = (ILNumerics.complex)processedData[i];
   //get frequency at i
   double freq = Common.sampleRate * i / theLength; // sampleRate / 2!
   if (freq >= Common.lBound && freq <= Common.uBound) // amplify this region
   {
   c.real /= 2;
   }
   }
   ILArray<double> fData = ILMath.real(ILMath.ifft(processedData));
   double[] final = fData.Values.ToArray<double>();
   
   return final; //change this later
   // why not return ILArray<double>?
   }

    

   Regarding NaN: make sure, not to confuse byte / double datatypes! This would produce NaNs as well. Make sure, no NaNs are coming from buffer. Checks on NaN can be done with:
   
   find(isnan(fprocBuf)).Length > 0, or
   fprocBuf[find(isnan(fprocBuf))] = MachineParameterDouble.eps

    

5. • CommentAuthormlew
   • CommentTimeJul 13th 2010

    
   Funnily enough, the problem seemed to solve itself when I converted the byte array from the directsound buffer to a double array. I'm willing to guess that the bit of code that read a double array from the buffer introduced some conversion problems that ended up manifesting themselves as NaN.
   
   Thanks for the help!