def get_metrics(truth, estimates):
#Compute the SDR by bss_eval from museval library ver.4
truth = truth[np.newaxis, :, np.newaxis]
estimates = estimates[np.newaxis, :, np.newaxis]
sdr, isr, sir, sar, perm = bss_eval_images(truth, estimates)
#The function 'bss_eval_sources' is NOT recommended by documentation
#[Ref] J. Le Roux et.al., "SDR-half-baked or well done?" (2018)
#[URL] https://arxiv.org/pdf/1811.02508.pdf
#sdr, sir, sar, perm = bss_eval_sources(truth, estimates)
return sdr[0, 0], isr[0, 0], sir[0, 0], sar[0, 0], perm[0, 0]
def permute_metric(
name: str,
ref: np.ndarray,
est: np.ndarray,
compute_permutation: bool = False,
fs: Optional[int] = None) -> Union[float, Tuple[float, list]]:
"""
Computation of SiSNR/PESQ/STOI in permutation/non-permutation mode
Args:
name: metric name
ref: array, reference signal (N x S or S, ground truth)
est: array, enhanced/separated signal (N x S or S)
compute_permutation: return permutation order or not
fs: sample rate of the audio
"""
if name == "sisnr":
return _permute_eval(aps_sisnr,
ref,
est,
compute_permutation=compute_permutation,
fs=fs)
elif name == "pesq":
return _permute_eval(aps_pesq,
ref,
est,
compute_permutation=compute_permutation,
fs=fs)
elif name == "stoi":
return _permute_eval(aps_stoi,
ref,
est,
compute_permutation=compute_permutation,
fs=fs)
elif name == "sdr":
if ref.ndim == 1:
ref, est = ref[None, :], est[None, :]
sdr, _, _, _, ali = bss_eval_images(ref[..., None],
est[..., None],
compute_permutation=True)
if compute_permutation:
return sdr.mean(), ali[:, 0].tolist()
else:
return sdr[0, 0]
else:
raise ValueError(f"Unknown name of the metric: {name}")
#Call my function for audio pre-processing
data, Fs = pre_processing(data, Fs, down_sam)
#Call my function for getting STFT (complex STFT amplitude)
print("Original sound")
sf.write("./result/Original_sound.wav", data, Fs)
Y, Fs, freqs, times = get_STFT(data, Fs, frame_length, frame_shift)
#Call my function for updating NMF basements and weights
H, U, P, loss = get_complexNMF(Y, num_iter, num_base, loss_func, spa_order)
#Call my function for getting inverse STFT
X = get_rec(H, U, P)
rec_wav, Fs = get_invSTFT(X, Fs, frame_length, frame_shift)
rec_wav = rec_wav[: int(data.shape[0])] #inverse stft includes residual part due to zero padding
print("Reconstructed sound")
sf.write("./result/Reconstructed_sound.wav", rec_wav, Fs)
#Call my function for decomposing basements
get_decompose(H, U, P, num_base, Fs, frame_length, frame_shift)
#Compute the SDR by bss_eval from museval library ver.4
data = data[np.newaxis, :, np.newaxis]
rec_wav = rec_wav[np.newaxis, :, np.newaxis]
sdr, isr, sir, sar, perm = bss_eval_images(data, rec_wav)
#sdr, sir, sar, perm = bss_eval_sources(truth, data) #Not recommended by documentation
print("SDR: {:.3f} [dB]".format(sdr[0, 0]))
#Call my function for displaying graph
display_graph(Y, X, times, freqs, loss_func, num_iter)
#print("The weight vector: {}".format(np.sort(E_w)[::-1]))
X = E_h[:, valid_M] @ (E_w[valid_M, np.newaxis] * E_u[valid_M, :])
#In the case of power spectrogram
if spec_type == "pow":
Y = np.sqrt(Y)
X = np.sqrt(X)
#Phase recovery
Y = Y * np.exp(1j*arg)
X = X * np.exp(1j*arg)
#Call my function for getting inverse STFT
original_wav, Fs = get_invSTFT(Y, Fs, frame_length, frame_shift)
original_wav = original_wav[: int(data.shape[0])] #inverse stft includes residual part due to zero padding
sf.write("./log/original.wav", original_wav, Fs)
#Call my function for getting inverse STFT
rec_wav, Fs = get_invSTFT(X, Fs, frame_length, frame_shift)
rec_wav = rec_wav[: int(data.shape[0])] #inverse stft includes residual part due to zero padding
sf.write("./log/approximated.wav", rec_wav, Fs)
#Compute the SDR by bss_eval from museval library ver.4
original_wav = original_wav[np.newaxis, :, np.newaxis]
rec_wav = rec_wav[np.newaxis, :, np.newaxis]
sdr, isr, sir, sar, perm = bss_eval_images(original_wav, rec_wav)
#sdr, sir, sar, perm = bss_eval_sources(truth, data) #Not recommended by documentation
print("SDR: {:.3f} [dB]".format(sdr[0, 0]))
#Call my function for displaying graph
display_graph(Y, X, times, freqs, bound, M_list)