How can I convert a .wav file to a float array ( edit the float array e.g. adding two signals) and back to a .wav file without whitenoise

Question

I am trying to program an auralization via Ray-Tracing in processing. To edit a sample over the information from the Ray Tracer, i need to convert a .wav File (File-Format: PCM-signed,16bit,stereo,2 bytes/frame, little endian) to an Float Array.

I converted the audio via an audioInputStream and a DataInputStream, where I am loading the audio into an byte Array.

Then I convert the byte Array to a float array like this.

byte[] samples;
float[] audio_data = float(samples); 

When I convert the float Array back to a .wav File, I'm getting the sound of the original Audio-File.

But when I'm adding another Float Array to the Original signal and convert it back to a. wav file via the method above(even if I'm adding the same signal), i get a white noise signal instead of the wanted signal (I can hear the original signal under the white noise modulated, but very very silent).

I read about this problem before, that there can be problems by the conversion from the float array to a byte array. That's because float is a 32bit datatype and byte (in java) is only 16 bits and somehow the bytes get mixed together wrong so the white noise is the result. In Processing there is a data type with signed 16bit integers (named: "short") but i can't modify the amplitude anymore, because therefore i need float values, which i can't convert to short.

I also tried to handle the overflow (amplitude) in the float array by modulating the signal from 16 bit values (-32768/32767) to values from -1/1 and back again after mixing (adding) the signals. The result gave me white noise. When i added more than 2 signals it gaves me nothing (nothing to hear).

The concrete Problem I want to solve is to add many signals (more than 1000 with a decent delay to create a kind of reverbation) in the form of float Arrays. Then I want to combine them to one Float Array that i want to save as an audio file without white noise.

I hope you guys can help me.

Answer 1

If you have true PCM data points, there should be no problem using simple addition. The only issue is that on rare occasions (assuming your audio is not too hot to begin with) the values will go out of range. This will tend create a harsh distortion, not white noise. The fact that you are getting white noise suggests to me that maybe you are not converting your PCM sums back to bytes correctly for the format that you are outputting.

Here is some code I use in AudioCue to convert PCM back to bytes. The format is assumed to be 16-bit, 44100 fps, stereo, little-endian. I'm working with PCM as normalized floats. This algorithm does the conversion for a buffer's worth of data at a time.

for (int i = 0, n = buffer.length; i < n; i++)
    {
        buffer[i] *= 32767;

        audioBytes[i*2] = (byte) buffer[i];
        audioBytes[i*2 + 1] = (byte)((int)buffer[i] >> 8 );
    }

Sometimes, a function like Math.min(Math.max(audioval, -1), 1) or Math.min(Math.max(audioval, -32767), 32767) is used to keep the values in range. More sophisticated limiters or compressor algorithms will scale the volume to fit. But still, if this is not handled, the result should be distortion, not white noise.

If the error is happening at another stage, we will need to see more of your code.

All this said, I wish you luck with the 1000-point echo array reverb. I hadn't heard of this approach working. Maybe there are processors that can handle the computational load now? (Are you trying to do this in real time?) My only success with coding real-time reverberation has been to use the Schroeder method, plugging the structure and values from the CCMRA Freeberb , working off of code from Craig Lindley's now ancient (copyright 2001) book "Digital Audio with Java". Most of that book deals with obsolete GUI code (pre-Swing!), but the code he gives for AllPass and Comb filters is still valid.

I recall when I was working on this that I tracked down references a better reverb to try and code, but I would have to do some real digging to try and find my notes. I was feeling over my head at the time, as the algorithm was presented via block diagrams not coding details or even pseudo-code. Would like to work on this again though and get a better reverb than the Shroeder-type to work. The Schoeder was passable for sounds that were not too percussive.

Getting a solution for real-time ray tracing would be a valuable accomplishment. Many applications in AR/VR and games.