clip spectrum when substracting noise in saliency computation

compute_salience_SHR.py

@@ -6,13 +6,13 @@ import librosa
 np.seterr(divide = 'ignore')
 
 parser = argparse.ArgumentParser()
-parser.add_argument('specie', type=str, help="Species to treat specifically", default=None)
+parser.add_argument('specie', type=str, help="Species to treat specifically", default='all')
 args = parser.parse_args()
 
-for specie in species if args.specie is None else [args.specie]:
-    wavpath, FS, nfft, downsample = species[specie].values()
+for specie in species if args.specie == 'all' else [args.specie]:
+    wavpath, FS, nfft, downsample, step = species[specie].values()
     Hz2bin = lambda f: np.floor(f / FS * nfft).astype(int) # f in _preds.csv are already downsamples
-    dt = nfft / 8 # winsize / 8
+    dt = nfft *step
     hann = np.hanning(nfft)
     get_spectrum = lambda x : 10 * np.log10(np.abs(np.fft.rfft(hann * x))**2 / (nfft/FS*1.5) +1) +1e-10 # density scaling, log1p 
 
@@ -35,7 +35,7 @@ for specie in species if args.specie is None else [args.specie]:
             if FS*r.time < nfft//2 or FS*r.time > len(sig) - nfft//2 or (sig[int(FS*r.time) - nfft//2 : int(FS*r.time) + nfft//2] == 0).all():
                 continue
             spec = get_spectrum(sig[int(FS*r.time) - nfft//2 : int(FS*r.time) + nfft//2])
-            spec -= noise
+            spec = np.clip(spec-noise, 0, 10)
             f0 = r.annot
 
             df.loc[i, 'salience'] = sum(spec[Hz2bin(f0*2**(-1/12)):Hz2bin(f0*2**(1/12))+1]) / sum(spec[Hz2bin(f0*2**(-6/12)):Hz2bin(f0*2**(6/12))+1])