def validate(self, name):
try:
noisy, _ = next(self.val_iter)
except:
self.val_iter = iter(self.train_dataloader)
noisy, _ = next(self.val_iter)
audio_utils.write_wav(
os.path.join(self.output_dir, "noisy_{}.wav".format(name)), noisy)
wandb.log({
"noisy": [
wandb.Audio(os.path.join(self.output_dir,
"noisy_{}.wav".format(name)),
caption="noisy",
sample_rate=16000)
]
})
with torch.no_grad():
m = noisy.mean()
noisy = (noisy - m)
noisy = torch.tensor([noisy]).to(self.device)
out = self.model(noisy)
clean = np.squeeze(out.cpu().detach().numpy())
clean = clean + m
audio_utils.write_wav(
os.path.join(self.output_dir, "clean_{}.wav".format(name)), clean)
wandb.log({
"clean": [
wandb.Audio(os.path.join(self.output_dir,
"clean_{}.wav".format(name)),
caption="clean",
sample_rate=16000)
]
})
torch.save(self.model.state_dict(),
os.path.join(self.output_dir, "{}.pth".format(name)))
def valid(self, step):
self.model.eval()
index = random.randint(0, len(self.noisy_files)-1)
noisy = dataloader.read_wav(self.noisy_files[index])
write_wav(os.path.join(self.output_dir, "noisy_{}.wav".format(step)), noisy)
wandb.log({"noisy": [wandb.Audio(os.path.join(self.output_dir, "noisy_{}.wav".format(step)), caption="noisy", sample_rate=self.sr)]})
noisy = torch.tensor([noisy], dtype=torch.float32, device=self.device)
with torch.no_grad():
clean, noise = self.model(noisy)
clean = np.squeeze(clean.cpu().detach().numpy())
write_wav(os.path.join(self.output_dir, "clean_{}.wav".format(step)), clean, self.sr)
wandb.log({"clean": [wandb.Audio(os.path.join(self.output_dir, "clean_{}.wav".format(step)), caption="clean", sample_rate=self.sr)]})
noise = np.squeeze(noise.cpu().detach().numpy())
write_wav(os.path.join(self.output_dir, "noise_{}.wav".format(step)), noise, self.sr)
wandb.log({"noise": [wandb.Audio(os.path.join(self.output_dir, "noise_{}.wav".format(step)), caption="noise", sample_rate=self.sr)]})
def compute_feature(bg_batch,
hm_batch,
_n_filt=N_FILT,
_winlen=WINLEN,
_winstep=WINLEN // 2,
to_write=False,
_winfunc=lambda x: np.ones((x, ))):
nfft = int(SAMPLING_RATE * _winlen)
bg_mfcc, bg_mfcc_energy = mfcc(bg_batch,
winlen=_winlen,
winstep=_winstep,
numcep=_n_filt,
nfft=nfft,
nfilt=_n_filt,
preemph=0,
cb=compute_gain,
ceplifter=0,
appendEnergy=False,
winfunc=_winfunc)
_, hm_mfcc_energy = mfcc(hm_batch,
winlen=_winlen,
winstep=_winstep,
numcep=_n_filt,
nfft=nfft,
nfilt=_n_filt,
preemph=0,
cb=compute_gain,
ceplifter=0,
appendEnergy=False,
winfunc=_winfunc)
# Ideal Ratio Mask applied signal
estimate = get_estimate(bg_batch,
np.clip(np.sqrt((hm_mfcc_energy / bg_mfcc_energy)),
0, 1),
_winlen=_winlen,
_winfunc=_winfunc,
_winstep=_winstep)
if to_write:
print('tes')
write_wav('mixed.wav', bg_batch.astype(np.float32))
write_wav('original.wav', hm_batch.astype(np.float32))
write_wav('ideal.wav', estimate.astype(np.float32))
return bg_mfcc
elif os.path.isfile(inp):
noisy_files = [inp]
else:
raise Exception("given input file/directory does not exist please check path")
if len(noisy_files) == 0:
raise Exception("folder does not contain files of format {}".format(args.format))
time_counter = []
for f in noisy_files:
audio = read_wav(f, sr)
start = time.time()
with torch.no_grad():
raw = torch.tensor([audio], dtype=torch.float32, device=device)
out = denoiser(raw)
out = np.squeeze(out.detach().cpu().numpy())
out = out
t = time.time() - start
time_counter.append(t)
print("processed file {} in {:.2f} secs".format(os.path.basename(f), t))
write_wav(os.path.join(audio_out, os.path.basename(f).split(".")[0] + "_clean.wav"), out, sr)
if args.asr:
transcript = ASR(f)
with open(os.path.join(trans_out, os.path.basename(f).split(".")[0] + ".json"), 'w') as outfile:
json.dump(transcript, outfile)
print("Time stats: min: {:.2f} mean: {:.2f} max: {:.2f}".format(min(time_counter), sum(time_counter)/len(time_counter), max(time_counter)))