d8ab8131e19498c2b9bed8d218e4e46eedc7cf21,librosa/tf_agc.py,,tf_agc,#,106

Before Change

            if f2a is None: 
                // initialize the mel filter bank after grabbing the first frame
                f2a = melfb(sample_rate, len(frame), num_frequency_bands, mel_filter_width)
                f2a = f2a[:,:(round(len(frame)/2) + 1)]

                n   = f2a.shape[0]
                

After Change

        // Iterate over frames
        for frame in frame_iterator:
            frame = numpy.frombuffer(frame, "h")

            if f2a is None: 
                // initialize the mel filter bank after grabbing the first frame

                f2a = melfb(sample_rate, len(frame), num_frequency_bands, mel_filter_width)
//                 f2a = f2a[:,:(round(len(frame)/2) + 1)]
                

                //% map back to FFT grid, flatten bark loop gain
                //sf2a = sum(f2a);

                normalize_f2a                       = numpy.sum(f2a, axis=0)
                normalize_f2a[normalize_f2a == 0]   = 1.0
                normalize_f2a                       = 1.0 / normalize_f2a

                // initialze the state vector
                state   = numpy.zeros( (num_frequency_bands, 1) )[0]

                pass

            // FFT each frame
            D = scipy.fft(frame)

            // multiply by f2a
            audiogram = numpy.dot(f2a, numpy.abs(D))

            //// DPWE
            //             state = max([alpha*state,audgram(:,i)],[],2);
            //             fbg(:,i) = state;
            // ...
            //
            state = numpy.maximum(alpha * state, audiogram)

            //E = diag(1./(sf2a+(sf2a==0))) * f2a" * fbg;
            E   = normalize_f2a * numpy.dot(f2a.T, state);

Instances