Audio data normalization range (values)

ajithkjajith · February 4, 2024, 4:43pm

Hi
Question/Issue:
I have specific question on the audio data that is used to visualize in the edge impulse project.
The amplitude of the original data will range between -1 to 1, but when I see in the visualization of the audio data that is upload. it ranges in 1000/10000 values.

How are these normalized?
How are these varying?
Also Can I know what is the range that it is scaled to bcoz input in the live classification or inference is also ranging in 1000/10000?

Eoin · February 22, 2024, 11:01am

Hi @ajithkjajith

The scaling factor applied either during feature extraction or as a post-processing step to ensure that the feature values are within a suitable range for the learn block used for classification or inference. (which is why you also see this on device)

github.com

edgeimpulse/processing-blocks/blob/d00b59a787caca86760a40b0cc615b13a65ba1ef/mfe/dsp.py#L56


      
          features = []
          graphs = []
          
          
width = 0
          height = 0
          
          
for ax in range(0, len(axes)):
              signal = raw_data[:,ax]
          
          
    if implementation_version >= 3:
                  # Rescale to [-1, 1] and add preemphasis
                  signal = (signal / 2**15).astype(np.float32)
                  signal = speechpy.processing.preemphasis(signal, cof=0.98, shift=1)
          
          
    # check for too many frames
              numframes, _, __ = speechpy.processing.calculate_number_of_frames(
                  signal,
                  implementation_version=implementation_version,
                  sampling_frequency=fs,
                  frame_length=frame_length,
                  frame_stride=frame_stride,

There are a number of parameters that can influence the scaling:

Hopefully that answers your question, if you have any further I may need to loop in the DSP team

Best

Eoin