def eval(ref_name, enh_name, nsy_name, results):
try:
utt_id = ref_name.split('/')[-1]
ref, sr = audioread(ref_name)
enh, sr = audioread(enh_name)
nsy, sr = audioread(nsy_name)
enh_len = enh.shape[0]
ref_len = ref.shape[0]
if enh_len > ref_len:
enh = enh[:ref_len]
else:
ref = ref[:enh_len]
nsy = nsy[:enh_len]
ref_score = pesq(ref, nsy, sr)
enh_score = pesq(ref, enh, sr)
ref_stoi = stoi(ref, nsy, sr, extended=False)
enh_stoi = stoi(ref, enh, sr, extended=False)
ref_sisdr = si_sdr(nsy, ref)
enh_sisdr = si_sdr(enh, ref)
ref_sdr = sdr(nsy, ref)
enh_sdr = sdr(enh, ref)
except Exception as e:
print(e)
results.append([
utt_id, {
'pesq': [ref_score, enh_score],
'stoi': [ref_stoi, enh_stoi],
'si_sdr': [ref_sisdr, enh_sisdr],
'sdr': [ref_sdr, enh_sdr],
}
])
def eval(self, audio_est, audio_ref):
x_est, fs_est = sf.read(audio_est)
x_ref, fs_ref = sf.read(audio_ref)
# align
len_x = np.min([len(x_est), len(x_ref)])
x_est = x_est[:len_x]
x_ref = x_ref[:len_x]
# x_ref = x_ref / np.max(np.abs(x_ref))
if fs_est != fs_ref:
raise ValueError(
'Sampling rate is different amon estimated audio and reference audio'
)
if self.metric == 'rmse':
return compute_rmse(x_est, x_ref)
elif self.metric == 'pesq':
return pesq(x_ref, x_est, fs_est)
elif self.metric == 'stoi':
return stoi(x_ref, x_est, fs_est, extended=False)
elif self.metric == 'estoi':
return stoi(x_ref, x_est, fs_est, extended=True)
elif self.metric == 'all':
score_rmse = compute_rmse(x_est, x_ref)
score_pesq = pesq(x_ref, x_est, fs_est)
score_stoi = stoi(x_ref, x_est, fs_est, extended=False)
return score_rmse, score_pesq, score_stoi
else:
raise ValueError(
'Evaluation only support: rmse, pesq, (e)stoi, all')
def calc_true_pesq(self, x, y, s, mask, fs=16000):
scores = []
for i in range(x.shape[0]):
ind = int(torch.sum(mask[i]))
x_i = x[i][:, :ind]
y_i = y[i][:, :ind]
s_i = s[i][:, :ind]
x_i = torch.istft(x_i,
n_fft=512,
win_length=512,
hop_length=128,
normalized=True).detach().cpu().numpy()
y_i = torch.istft(y_i,
n_fft=512,
win_length=512,
hop_length=128,
normalized=True).detach().cpu().numpy()
s_i = torch.istft(s_i,
n_fft=512,
win_length=512,
hop_length=128,
normalized=True).detach().cpu().numpy()
score_x = pesq(s_i, x_i, fs)
score_y = pesq(s_i, y_i, fs)
score_s = pesq(s_i, s_i, fs)
del x_i
del y_i
del s_i
scores.append([score_x, score_y, score_s])
return torch.tensor(scores)
def eval_pesq(predicted_path,
noisy_test,
clean_test,
out_path,
img_path='/u/anakuzne/data/snr0_test_img/',
fs=16000):
'''
Params:
predicted: stfts predicted by the model
noisy test: stfts of noise mixture
clean test: clean unmixed test signals
'''
noisy_ref = os.listdir(noisy_test) #.wav
clean_ref = collect_paths(clean_test) #.wav
predicted = os.listdir(predicted_path) #.npy
imag = os.listdir(img_path) #.npy
scores_clean_ref = []
scores_noisy_ref = []
print('Calculating PESQ...')
for p in tqdm(clean_ref):
clean_r, sr = librosa.load(p, mono=True)
clean_r = librosa.core.resample(clean_r, sr, 16000)
fname = p.split('/')[-2] + '_' + p.split('/')[-1].split(
'.')[0] + '.npy'
ind = predicted.index(fname)
pred = np.load(predicted_path + predicted[ind])
ind = imag.index(fname)
imag_num = pad(np.load(img_path + imag[ind]), 1339)
pred = pred + imag_num
pred = librosa.istft(pred, hop_length=256, win_length=512)
pred = pred[:clean_r.shape[0]]
##Compare to clean signal
score_clean = pesq(clean_r, pred, fs)
scores_clean_ref.append(score_clean)
#Compare to degraded signal
ind = noisy_ref.index(fname)
pred = np.load(predicted_path + predicted[ind])
noisy_r = np.load(noisy_test + noisy_ref[ind])
noisy_r = librosa.istft(noisy_r, hop_length=256, win_length=512)
pred = librosa.istft(pred, hop_length=256, win_length=512)
pred = pred[:noisy_r.shape[0]]
score_noisy = pesq(noisy_r, pred, fs)
scores_noisy_ref.append(score_noisy)
wav_names = [n.split('/')[-1] for n in clean_ref]
data = {
'fname': wav_names,
'PESQ_clean_ref': scores_clean_ref,
'PESQ_noisy_ref': scores_noisy_ref
}
df = pd.DataFrame(data,
columns=['fname', 'PESQ_clean_ref', 'PESQ_noisy_ref'])
df.to_csv(out_path + 'PESQ.csv')
def getPESQ(name, clean_speech_dir, noise_num):
clean_waves_dir = 'exp/data_for_ac/' + clean_speech_dir
clean_list = os.listdir(clean_waves_dir)
clean_dir_list = [
os.path.join(clean_waves_dir, clean_file) for clean_file in clean_list
]
clean_dir_list.sort()
enhanced_waves_dir = 'exp/data_for_ac/' + name + '/enhanced_wav'
enhanced_list = os.listdir(enhanced_waves_dir)
enhanced_dir_list = [
os.path.join(enhanced_waves_dir, enhanced_file)
for enhanced_file in enhanced_list
]
enhanced_dir_list.sort()
mixed_waves_dir = 'exp/data_for_ac/' + name + '/mixed_wav'
mixed_list = os.listdir(mixed_waves_dir)
mixed_dir_list = [
os.path.join(mixed_waves_dir, mixed_file) for mixed_file in mixed_list
]
mixed_dir_list.sort()
clean_dir_list_long = []
for clean_dir in clean_dir_list:
for i in range(noise_num):
clean_dir_list_long.append(clean_dir)
avg_score_raw = 0.0
avg_score_en = 0.0
avg_score_imp = 0.0
i = 0
for clean_wave, enhanced_wave, mixed_wave in zip(clean_dir_list_long,
enhanced_dir_list,
mixed_dir_list):
ref, sr = sf.read(clean_wave)
mixed, sr = sf.read(mixed_wave)
enhanced, sr = sf.read(enhanced_wave)
# spec = spectrum_tool.magnitude_spectrum_librosa_stft(enhanced,512,256)
# angle = spectrum_tool.phase_spectrum_librosa_stft(enhanced,512,256)
# spec = spec ** 1.3
# enhanced = spectrum_tool.librosa_istft(spec*np.exp(angle*1j),512,256)
# score_raw = pesq(ref/np.max(np.abs(ref)), mixed/np.max(np.abs(mixed)), sr)
# score_en = pesq(ref/np.max(np.abs(ref)), enhanced/np.max(np.abs(enhanced)), sr)
score_raw = pesq(ref, mixed, sr)
score_en = pesq(ref, enhanced, sr)
print(
str(i % noise_num + 1) + "_score_raw, score_en:", score_raw,
score_en)
i += 1
avg_score_raw += score_raw
avg_score_en += score_en
avg_score_imp += (score_en - score_raw)
avg_score_raw /= len(clean_dir_list_long)
avg_score_en /= len(clean_dir_list_long)
avg_score_imp /= len(clean_dir_list_long)
print('avg_score_raw: %f,\navg_score_en: %f,\nimp: %f' %
(avg_score_raw, avg_score_en, avg_score_imp))
def separate_sample(sess, model, config, mix, c1, c2):
batch_size = config['training']['batch_size']
n_channel = config['training']['n_output']
n_speaker = config['training']['n_speaker']
stride = config['model']['hop']
num_output_samples = mix.shape[0]
num_fragments = int(np.ceil(num_output_samples / model.input_length))
target_field_length = model.input_length
target_padding = 0
num_batches = int(np.ceil(num_fragments / batch_size))
output = [[] for _ in range(n_channel)]
num_pad_values = 0
fragment_i = 0
# pad input mixture to 10x, since stride is 10
num_pad_values = stride - mix.shape[0] % stride
mix = np.pad(mix, (0, num_pad_values), mode='constant', constant_values=0)
output, = sess.run(fetches=[model.data_out],
feed_dict={model.mix_input: np.expand_dims(mix, 0)})[0]
output = np.array(output)
if num_pad_values != 0:
output = output[:, :-num_pad_values]
mix = mix[:-num_pad_values]
clean_wav = np.array([c1, c2])
perms = np.array(list(itertools.permutations(range(n_channel), n_speaker)))
perms_onehot = (np.arange(perms.max() + 1) == perms[..., None]).astype(int)
cross_sdr = signal_to_distortion_ratio(
np.expand_dims(np.array([c1, c2]), 1), np.expand_dims(output, 0))
loss_sets = np.einsum('ij,pij->p', cross_sdr, perms_onehot)
best_perm = perms[np.argmax(loss_sets)]
pit_output = output[best_perm]
# SDR
_sdr, _sir, _sar, _perm = mir_eval.separation.bss_eval_sources(
clean_wav, pit_output)
# SISNR
clean_wav_norm = clean_wav - np.mean(clean_wav, axis=-1, keepdims=True)
pit_output_norm = pit_output - np.mean(pit_output, axis=-1, keepdims=True)
_sisnr = signal_to_distortion_ratio(clean_wav_norm, pit_output_norm)
# PESQ
_pesq = [
pesq(clean_wav[0], pit_output[0], 8000),
pesq(clean_wav[1], pit_output[1], 8000)
]
perm_output = np.expand_dims(best_perm, -1).tolist()
return _sdr, _sisnr, _pesq, perm_output
def evaluation(ref, est, mix):
"""
Wrapper function for evaluating the output of a NN. Metrics are PESQ and STOI
:param ref: Path to the original (reference point) file.
:param est: Path to the estimated file.
:return: Prints in stdout PESQ and STOI metric values.
"""
file_ref = ref
file_est = est
file_mix = mix
reference_sources, sr_r = librosa.load(file_ref, sr = None)
estimated_sources, sr_e = librosa.load(file_est, sr = None)
mix_sources, sr_m = librosa.load(file_mix, sr = None)
if sr_r != 16000 or sr_e != 16000 or sr_m != 16000:
print("\nResampling at 16k...")
ref_16k = librosa.resample(reference_sources, sr_r, 16000)
est_16k = librosa.resample(estimated_sources, sr_e, 16000)
mix_16k = librosa.resample(mix_sources, sr_e, 16000)
else:
ref_16k = reference_sources
est_16k = estimated_sources
mix_16k = mix_sources
pesq_score = round(pesq(ref_16k, est_16k, 16000), 3)
stoi_score = round(stoi(ref_16k, est_16k, sr_r, extended=False), 2)
estoi_score = round(stoi(ref_16k, est_16k, sr_r, extended=True), 2)
ssr_score = round(SSR(est_16k, mix_16k), 3)
print("PESQ\t STOI\t eSTOI\t SSR")
print(pesq_score,"\t",stoi_score,"\t",estoi_score,"\t",ssr_score)
def PESQ(ref_wav, deg_wav):
# reference wav
# degraded wav
sr = 16000
# tfl = tempfile.NamedTemporaryFile()
# ref_tfl = tfl.name + '_ref.wav'
# deg_tfl = tfl.name + '_deg.wav'
#
# # if ref_wav.max() <= 1:
# # ref_wav = np.array(denormalize_wave_minmax(ref_wav), dtype=np.int16)
# # if deg_wav.max() <= 1:
# # deg_wav = np.array(denormalize_wave_minmax(deg_wav), dtype=np.int16)
#
# # wavfile.write(ref_tfl, 16000, ref_wav)
# # wavfile.write(deg_tfl, 16000, deg_wav)
# sf.write(ref_tfl, ref_wav, sr, subtype='PCM_16')
# sf.write(deg_tfl, deg_wav, sr, subtype='PCM_16')
# curr_dir = os.getcwd()
# Write both to tmp files and then eval with pesqmain
# try:
# p = run(['pesqmain'.format(curr_dir),
# ref_tfl, deg_tfl, '+'+str(sr), '+wb'],
# stdout=PIPE,
# encoding='ascii')
# res_line = p.stdout.split('\n')[-2]
# results = re.split('\s+', res_line)
# return results[-1]
# except FileNotFoundError:
# print('pesqmain not found! Please add it your PATH')
score = pesq(ref_wav, deg_wav, sr)
return score
def calc_metrics(loader, actor, device):
pesq_all = []
stoi_all = []
for batch in loader:
x = batch["noisy"].unsqueeze(1).to(device)
t = batch["clean"].unsqueeze(1).to(device)
m = batch["mask"].to(device)
out_r, out_i = actor(x)
out_r = torch.transpose(out_r, 1, 2)
out_i = torch.transpose(out_i, 1, 2)
y = predict(x.squeeze(1), (out_r, out_i))
t = t.squeeze()
m = m.squeeze()
#print("Y:", y.shape)
#source, targets, preds = inverse(t, y, m, x)
targets, preds = inverse(t, y, m, x)
for j in range(len(targets)):
curr_pesq = pesq(targets[j].detach().cpu().numpy(),
preds[j].detach().cpu().numpy(), 16000)
curr_stoi = stoi(targets[j].detach().cpu().numpy(),
preds[j].detach().cpu().numpy(), 16000)
pesq_all.append(curr_pesq)
stoi_all.append(curr_stoi)
PESQ = torch.mean(torch.tensor(pesq_all))
STOI = torch.mean(torch.tensor(stoi_all))
return PESQ, STOI
def cal_score(clean, enhanced):
clean = clean / abs(clean).max()
enhanced = enhanced / abs(enhanced).max()
s_stoi = stoi(clean, enhanced, 16000)
s_pesq = pesq(clean, enhanced, 16000)
return round(s_pesq, 5), round(s_stoi, 5)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--results_dir', type=str, required=True)
parser.add_argument('--audio_sampling_rate', type=int, default=16000)
args = parser.parse_args()
audio1, _ = librosa.load(os.path.join(args.results_dir,
'audio1_separated.wav'),
sr=args.audio_sampling_rate)
audio2, _ = librosa.load(os.path.join(args.results_dir,
'audio2_separated.wav'),
sr=args.audio_sampling_rate)
audio1_gt, _ = librosa.load(os.path.join(args.results_dir, 'audio1.wav'),
sr=args.audio_sampling_rate)
audio2_gt, _ = librosa.load(os.path.join(args.results_dir, 'audio2.wav'),
sr=args.audio_sampling_rate)
audio_mix, _ = librosa.load(os.path.join(args.results_dir,
'audio_mixed.wav'),
sr=args.audio_sampling_rate)
# SDR, SIR, SAR
sdr, sir, sar = getSeparationMetrics(audio1, audio2, audio1_gt, audio2_gt)
sdr_mixed, _, _ = getSeparationMetrics(audio_mix, audio_mix, audio1_gt,
audio2_gt)
# PESQ
pesq_score1 = pesq(audio1, audio1_gt, args.audio_sampling_rate)
pesq_score2 = pesq(audio2, audio2_gt, args.audio_sampling_rate)
pesq_score = (pesq_score1 + pesq_score2) / 2
# STOI
stoi_score1 = stoi(audio1_gt,
audio1,
args.audio_sampling_rate,
extended=False)
stoi_score2 = stoi(audio2_gt,
audio2,
args.audio_sampling_rate,
extended=False)
stoi_score = (stoi_score1 + stoi_score2) / 2
output_file = open(os.path.join(args.results_dir, 'eval.txt'), 'w')
output_file.write(
"%3f %3f %3f %3f %3f %3f %3f" %
(sdr, sdr_mixed, sdr - sdr_mixed, sir, sar, pesq_score, stoi_score))
output_file.close()
def forward(self, clean, enhanced):
assert len(clean) == len(enhanced)
clean, enhanced = np.array(clean), np.array(enhanced)
scores = []
for c, e in zip(clean, enhanced):
q = pesq(c.detach().cpu(), e.detach().cpu(), self.sr)
q = (q + 0.5) / (4.5 + 0.5) # (q-min)/(max-min)
scores.append(q)
return torch.tensor(scores, device=self._device)
def compute_PESQ(clean_signal, noisy_signal, sr=16000):
"""计算PESQ
Args:
clean_signal:纯净语音信号
noisy_signal:带噪语音信号
sr: 采样率
"""
return pesq(clean_signal, noisy_signal, sr)
def calculate_pesq(real, fake, sr):
#PESQ only work with sampling rates 8000 or 16000, so we must resample
try:
if sr != 8000:
real = librosa.resample(real, sr, 8000)
fake = librosa.resample(fake, sr, 8000)
return pesq(real, fake, 8000)
except:
return 0.0
def inference(clean_path, noisy_path, model_path, out_path):
device = torch.device("cuda:1")
model = Actor()
model = nn.DataParallel(model, device_ids=[1, 2])
model.load_state_dict(torch.load(model_path + 'actor_best.pth'))
model = model.to(device)
dataset = Data(clean_path, noisy_path, mode='Test')
loader = data.DataLoader(dataset,
batch_size=32,
shuffle=False,
collate_fn=collate_custom)
fnames = os.listdir(noisy_path)
print("Num files:", len(fnames))
pesq_all = []
stoi_all = []
fcount = 0
for batch in tqdm(loader):
x = batch["noisy"].unsqueeze(1).to(device)
t = batch["clean"].unsqueeze(1).to(device)
m = batch["mask"].to(device)
out_r, out_i = model(x)
out_r = torch.transpose(out_r, 1, 2)
out_i = torch.transpose(out_i, 1, 2)
y = predict(x.squeeze(1), (out_r, out_i))
t = t.squeeze()
m = m.squeeze()
x = x.squeeze()
source, targets, preds = inverse(t, y, m, x)
for j in range(len(targets)):
t_j = targets[j].detach().cpu().numpy()
p_j = preds[j].detach().cpu().numpy()
p_j = 10 * (p_j / np.linalg.norm(p_j))
curr_pesq = pesq(t_j, p_j, 16000)
curr_stoi = stoi(t_j, p_j, 16000)
pesq_all.append(curr_pesq)
stoi_all.append(curr_stoi)
try:
sf.write(os.path.join(out_path, fnames[fcount]), p_j, 16000)
except IndexError:
print("Fcount:", fcount, len(fnames))
fcount += 1
PESQ = torch.mean(torch.tensor(pesq_all))
STOI = torch.mean(torch.tensor(stoi_all))
print("PESQ: ", PESQ, "STOI: ", STOI)
with open(os.path.join(model_path, 'test_scores.txt'), 'w') as fo:
fo.write("Avg PESQ: " + str(float(PESQ)) + " Avg STOI: " +
str(float(STOI)))
def __call__(self, a, b):
"""
:param a: 时域信号
:param b: 时域信号
:return:
"""
assert len(a.shape) == 1
assert len(a) == len(b)
score = pesq(a, b, self.sr)
return score
def pesq_by_inout(in_file, out_file):
sr, ref = read(in_file)
sr, deg = read(out_file)
if len(ref.shape) >= 2:
ref = ref[:, 0]
if len(deg.shape) >= 2:
deg = deg[:, 0]
ref = np.pad(ref, (0, len(deg) - len(ref)), "constant", constant_values=0)
return pesq(ref, deg, sr)
def compute_PESQ(clean_signal, noisy_signal, sr=16000):
"""
使用 pypesq 计算 pesq 评价指标。
Notes：
pypesq 是 PESQ 官方代码（C 语言）的 wrapper。官方代码在某些语音上会报错，而其报错时直接调用了 exit() 函数，直接导致 Python 脚本停止运行，亦无法捕获异常，实在过于粗鲁。
为了防止 Python 脚本被打断，这里使用子进程执行 PESQ 评价指标的计算，设置子进程的超时。
设置子进程超时的方法：https://stackoverflow.com/a/29378611
Returns:
当语音可以计算 pesq score 时，返回 pesq score，否则返回 -1
"""
return pesq(clean_signal, noisy_signal, sr)
def pesq_metric(y_hat, y_true):
with torch.no_grad():
y_hat = y_hat.cpu().data.numpy()
y = y_true.cpu().data.numpy()
sum_pesq = 0
for i in range(len(y)):
sum_pesq += pesq(y[i], y_hat[i], SAMPLE_RATE)
sum_pesq /= len(y)
return sum_pesq
def eval(ref_name, enh_name, nsy_name, results):
try:
utt_id = ref_name.split('/')[-1]
ref, sr = audioread(ref_name)
enh, sr = audioread(enh_name)
nsy, sr = audioread(nsy_name)
ref_score = pesq(ref, nsy, sr)
enh_score = pesq(ref, enh, sr)
ref_stoi = stoi(ref, nsy, sr, extended=False)
enh_stoi = stoi(ref, enh, sr, extended=False)
ref_sdr = si_sdr(nsy, ref)
enh_sdr = si_sdr(enh, ref)
except Exception as e:
print(e)
results.append([utt_id,
{'pesq':[ref_score, enh_score],
'stoi':[ref_stoi,enh_stoi],
'si_sdr':[ref_sdr, enh_sdr]
}])
def pesq_score(self, ref_audio: 'Audio') -> float:
"""
PESQ score for speech audio.
Args:
ref_audio: Reference audio of type `Audio`.
Returns:
float -- PESQ score.
"""
assert ref_audio.sr == self.sr, 'Sample rate should be same'
return pypesq.pesq(ref_audio.wave, self.wave, self.sr)
def pesq_metric(y_hat, bd):
# PESQ
with torch.no_grad():
y_hat = y_hat.cpu().numpy()
y = bd['y'].cpu().numpy() # target signal
sum = 0
for i in range(len(y)):
sum += pesq(y[i, 0], y_hat[i, 0], SAMPLE_RATE)
sum /= len(y)
return torch.tensor(sum)
def _eval(batch,
metrics,
including='output',
sample_rate=8000,
use_pypesq=False):
if use_pypesq:
metrics = [m for m in metrics if m != 'pesq']
has_estoi = False
if 'estoi' in metrics:
metrics = [m for m in metrics if m != 'estoi']
has_estoi = True
has_wer = False
if 'wer' in metrics:
metrics = [m for m in metrics if m != 'wer']
has_wer = True
mix = batch['mix']
clean = batch['clean']
estimate = batch['enh']
snr = batch['snr']
res = get_metrics(mix.numpy(),
clean.numpy(),
estimate.numpy(),
sample_rate=sample_rate,
metrics_list=metrics,
including=including)
if use_pypesq:
res['pesq'] = pesq(clean.flatten(), estimate.flatten(), sample_rate)
if has_estoi:
res['estoi'] = stoi(clean.flatten(),
estimate.flatten(),
sample_rate,
extended=True)
if has_wer:
res['wer'] = jiwer.wer(batch['clean_text'],
batch['transcription'],
truth_transform=_wer_trans,
hypothesis_transform=_wer_trans)
if including == 'input':
for m in metrics:
res[m] = res['input_' + m]
del res['input_' + m]
res['snr'] = snr[0].item()
return res
def generate_curriculum(clean_path, noisy_path, model_path):
fnames = os.listdir(clean_path)
d = {"fname": [], "pesq": []}
for fname in tqdm(fnames):
wav_noisy, sr = librosa.core.load(os.path.join(noisy_path, fname),
sr=16000)
wav_clean, sr = librosa.core.load(os.path.join(clean_path, fname),
sr=16000)
score = pesq(wav_clean, wav_noisy)
d["fname"].append(fname)
d["pesq"].append(score)
df = pd.DataFrame.from_dict(d)
df = df.sort_values(by=['pesq'])
df.to_csv(os.path.join(model_path, "train_sort.tsv"), sep='\t')
def loss_pesq(step, source, estimate_source):
source = source.cpu()
estimate_source = estimate_source.cpu()
score = 0
for i in range(source.shape[0]):
score = score + pesq(source[i], estimate_source[i], sr)
torchaudio.save(os.path.join(dns_home, 'predict_folder',
test[step * batch_size + i]),
estimate_source[i].unsqueeze(0),
sample_rate=16000)
return score / source.shape[0]
def cal_PESQi(src_ref, src_est, mix):
"""Calculate Source-to-Distortion Ratio improvement (SDRi).
NOTE: bss_eval_sources is very very slow.
Args:
src_ref: numpy.ndarray, [C, T]
src_est: numpy.ndarray, [C, T], reordered by best PIT permutation
mix: numpy.ndarray, [T]
Returns:
average_SDRi
"""
num = 0
new_pesq = 0
orig_pesq = 0
avg_PESQi = 0
for ref, est in zip(src_ref, src_est):
num = num + 1
new_pesq_out = pesq(ref, est, 8000)
new_pesq = new_pesq + new_pesq_out
orig_pesq_out = pesq(ref, mix, 8000)
orig_pesq = orig_pesq + orig_pesq_out
avg_PESQi = avg_PESQi + (new_pesq_out - orig_pesq_out)
return new_pesq / num, orig_pesq / num, avg_PESQi / num
def get_pesq(filepath_1, filepath_2):
'''
argument:
filepath_1: original clean .wav file path
filepath_2: generated clean .wav file path
return:
PESQ score
'''
clean_wave, clean_sampe_rate = librosa.load(filepath_1, sr=16000)
cleaned_wave, cleaned_sampe_rate = librosa.load(filepath_2, sr=16000)
signal_power = 20 * log10(np.mean(np.square(clean_wave)))
score = pesq(clean_wave, cleaned_wave, clean_sampe_rate)
return score, signal_power
def main():
parser = argparse.ArgumentParser(description='Calculate performance index')
parser.add_argument('--test_mix_folder',
default='../test-mix-2-babble',
type=str,
help='test-set-mix')
parser.add_argument('--test_clean_folder',
default='../test-clean-2-babble',
type=str,
help='test-set-clean')
parser.add_argument('--enhanced_folder',
default='../test-result',
type=str,
help='test-set-enhanced')
opt = parser.parse_args()
MIX_FOLDER = opt.test_mix_folder
CLEAN_FOLDER = opt.test_clean_folder
ENHANCED_FOLDER = opt.enhanced_folder
pesqs = []
stois = []
for cleanfile in os.listdir(CLEAN_FOLDER):
mixfile = cleanfile.replace('clean', 'mix')
enhancedfile = 'enhanced_' + mixfile
cleanfile = os.path.join(CLEAN_FOLDER, cleanfile)
mixfile = os.path.join(MIX_FOLDER, mixfile)
enhancedfile = os.path.join(ENHANCED_FOLDER, enhancedfile)
ref, sr1 = librosa.load(cleanfile, 16000)
#deg_mix, sr2 = librosa.load(mixfile, 16000)
deg_enh, sr3 = librosa.load(enhancedfile, 16000)
#pesq1 = pesq.pesq(ref, deg_mix)
pesq2 = pesq.pesq(ref, deg_enh[:len(ref)])
#print("pesq:", pesq1, " --> ", pesq2)
pesqs.append(pesq2)
#stoi1 = stoi(ref, deg_mix, fs_sig=16000)
stoi2 = stoi(ref, deg_enh[:len(ref)], fs_sig=16000)
#print("stoi:", stoi1, " --> ", stoi2)
stois.append(stoi2)
print('Epesq:', np.mean(pesqs), "Estoi:", np.mean(stois))
def cal_score(clean, enhanced):
clean = clean / abs(clean).max()
enhanced = enhanced / abs(enhanced).max()
s_stoi = stoi(clean, enhanced, 16000)
s_pesq = pesq(clean, enhanced, 16000)
return round(s_pesq, 5), round(s_stoi, 5)
#def get_filepaths(directory,folders='BabyCry.wav,cafeteria_babble.wav',ftype='.wav'):
# file_paths = []
# folders = folders.split(',')
# with open(directory, 'r') as f:
# for line in f:
# if str(line.split('/')[-3]) in folders:
# file_paths.append(line[:-1])
# return file_paths
'''
def sample_main():
sr = 16000
wav_list = ['data/1.wav', 'data/2.wav', 'data/3.wav', 'data/4.wav', 'data/5.wav']
src, mixed = mix_audio(wav_list)
from sklearn.decomposition import FastICA
W = FastICA(n_components=len(wav_list))
S = W.fit_transform(mixed)
## write audio into file
for i in range(len(wav_list)):
output_name = "result/%d.wav"%(i+1)
wavfile.write(output_name, sr, S[:, i])
"""
Caluculate sdr, sir, sar
"""
from mir_eval.separation import bss_eval_sources
sdr, sir, sar, _ = bss_eval_sources(S.T, src.T) ## shape = (channels, samples)
print("SDR: ", sdr) ## np.array(channels, )
print("SIR: ", sir)
print("SAR: ",sar)
from pypesq import pesq
pesq_score = pesq(src, S, fs=16000)
print("PESQ: ", pesq_score)
