def make_prediction(args):
model = load_model(args.model_fn,
custom_objects={
'Melspectrogram': Melspectrogram,
'Normalization2D': Normalization2D
})
wav_paths = glob('{}/**'.format('wavfiles'), recursive=True)
wav_paths = sorted([x for x in wav_paths if '.wav' in x])
classes = sorted(os.listdir('wavfiles'))
for wav_fn in wav_paths:
rate, wav = downsample_mono(wav_fn, args.sr)
mask, env = envelope(wav, rate, threshold=args.threshold)
clean_wav = wav[mask]
step = int(args.sr * args.dt)
batch = []
for i in range(0, clean_wav.shape[0], step):
sample = clean_wav[i:i + step]
sample = sample.reshape(1, -1)
if sample.shape[0] < step:
tmp = np.zeros(shape=(1, step), dtype=np.int16)
tmp[:, :sample.shape[1]] = sample.flatten()
sample = tmp
batch.append(sample)
X_batch = np.array(batch)
y_pred = model.predict(X_batch)
y_mean = np.mean(y_pred, axis=0)
y_pred = np.argmax(y_mean)
print(wav_fn, classes[y_pred])
def make_prediction(args):
model = load_model(args.model_fn,
custom_objects={
'STFT': STFT,
'Magnitude': Magnitude,
'ApplyFilterbank': ApplyFilterbank,
'MagnitudeToDecibel': MagnitudeToDecibel
})
wav_paths = glob('{}/**'.format(args.src_dir), recursive=True)
wav_paths = sorted(
[x.replace(os.sep, '/') for x in wav_paths if '.wav' in x])
classes = sorted(os.listdir(args.src_dir))
labels = [os.path.split(x)[0].split('/')[-1] for x in wav_paths]
le = LabelEncoder()
y_true = le.fit_transform(labels)
results = []
ylabelsResults = []
ytrueLab = []
for z, wav_fn in tqdm(enumerate(wav_paths), total=len(wav_paths)):
rate, wav = downsample_mono(wav_fn, args.sr)
mask, env = envelope(wav, rate, threshold=args.threshold)
clean_wav = wav[mask]
step = int(args.sr * args.dt)
batch = []
for i in range(0, clean_wav.shape[0], step):
sample = clean_wav[i:i + step]
sample = sample.reshape(-1, 1)
if sample.shape[0] < step:
tmp = np.zeros(shape=(step, 1), dtype=np.float32)
tmp[:sample.shape[0], :] = sample.flatten().reshape(-1, 1)
sample = tmp
batch.append(sample)
if (batch != []):
X_batch = np.array(batch, dtype=np.float32)
y_pred = model.predict(X_batch)
y_mean = np.mean(y_pred, axis=0)
y_pred = np.argmax(y_mean)
real_class = os.path.dirname(wav_fn).split('/')[-1]
print('Actual class: {}, Predicted class: {}'.format(
real_class, classes[y_pred]))
results.append(y_mean)
ylabelsResults.append(labels[z])
labenc = LabelEncoder()
ytrueLab = labenc.fit_transform(ylabelsResults)
np.save(os.path.join('logs', args.true_fn), np.array(ytrueLab))
np.save(os.path.join('logs', args.pred_fn), np.array(results))
def make_prediction(args):
model = load_model(args.model_fn,
custom_objects={
'Melspectrogram': Melspectrogram,
'Normalization2D': Normalization2D
})
wav_paths = glob('{}/**'.format('wavfiles'), recursive=True)
wav_paths = sorted(
[x.replace(os.sep, '/') for x in wav_paths if '.wav' in x])
classes = sorted(os.listdir('wavfiles'))
labels = [os.path.split(x)[0].split('/')[-1] for x in wav_paths]
le = LabelEncoder()
y_true = le.fit_transform(labels)
results = []
for z, wav_fn in tqdm(enumerate(wav_paths), total=len(wav_paths)):
try:
rate, wav = downsample_mono(src_fn, args.sr)
except:
print('Error with downsampling')
continue
mask, env = envelope(wav, rate, threshold=args.threshold)
clean_wav = wav[mask]
step = int(args.sr * args.dt)
batch = []
for i in range(0, clean_wav.shape[0], step):
sample = clean_wav[i:i + step]
sample = sample.reshape(1, -1)
if sample.shape[0] < step:
tmp = np.zeros(shape=(1, step), dtype=np.int16)
tmp[:, :sample.shape[1]] = sample.flatten()
sample = tmp
batch.append(sample)
X_batch = np.array(batch)
y_pred = model.predict(X_batch)
y_mean = np.mean(y_pred, axis=0)
y_pred = np.argmax(y_mean)
results.append(y_mean)
#np.save(os.path.join('logs', args.pred_fn), np.array(results))
return np.array(results)
def make_prediction(args):
#Welches feature soll benutzt werden?
model = load_model(args.model_fn,
custom_objects={
'Melspectrogram':
kapre.time_frequency.Melspectrogram,
'Normalization2D': Normalization2D
})
wav_paths = glob('{}/**'.format(args.src_dir), recursive=True)
wav_paths = sorted(
[x.replace(os.sep, '/') for x in wav_paths if '.wav' in x])
classes = sorted(os.listdir(args.src_dir))
labels = [os.path.split(x)[0].split('/')[-1] for x in wav_paths]
le = LabelEncoder()
y_true = le.fit_transform(labels)
results = []
for z, wav_fn in tqdm(enumerate(wav_paths), total=len(wav_paths)):
rate, wav = downsample_mono(wav_fn, args.sr)
mask, env = envelope(wav, rate, threshold=args.threshold)
clean_wav = wav[mask]
step = int(args.sr * args.dt)
batch = []
for i in range(0, clean_wav.shape[0], step):
sample = clean_wav[i:i + step]
sample = sample.reshape(1, -1)
if sample.shape[0] < step:
tmp = np.zeros(shape=(1, step), dtype=np.int16)
tmp[:, :sample.shape[1]] = sample.flatten()
sample = tmp
batch.append(sample)
X_batch = np.array(batch)
y_pred = model.predict(X_batch)
y_mean = np.mean(y_pred, axis=0)
y_pred = np.argmax(y_mean)
real_class = os.path.dirname(wav_fn).split('/')[-1]
print('Actual class: {}, Predicted class: {}'.format(
real_class, classes[y_pred]))
results.append(y_mean)
np.save(os.path.join('logs', args.pred_fn), np.array(results))
def make_prediction(model, args, address):
# print("Converting mp3 to wav... This may take a while ")
# convert(args.src_dir)
# wav_paths = glob('{}/**'.format(args.src_dir), recursive=True)
# wav_paths = sorted([x.replace(os.sep, '/') for x in wav_paths if '.wav' in x])
# classes = sorted(os.listdir(args.src_dir))
# labels = [os.path.split(x)[0].split('/')[-1] for x in wav_paths]
# le = LabelEncoder()
# y_true = le.fit_transform(labels)
# Convert bytes into wav file
results = []
wav_paths = [address] # only one file
ret_dict = {k: 0 for k in CLASSES.values()}
for z, wav_fn in tqdm(enumerate(wav_paths), total=len(wav_paths)):
rate, wav = downsample_mono(wav_fn, args.sr)
mask, env = envelope(wav, rate, threshold=args.threshold)
clean_wav = wav[mask]
step = int(args.sr * args.dt)
batch = []
for i in range(0, clean_wav.shape[0], step):
sample = clean_wav[i:i + step]
sample = sample.reshape(-1, 1)
if sample.shape[0] < step:
tmp = np.zeros(shape=(step, 1), dtype=np.float32)
tmp[:sample.shape[0], :] = sample.flatten().reshape(-1, 1)
sample = tmp
batch.append(sample)
X_batch = np.array(batch, dtype=np.float32)
y_pred = model.predict(X_batch)
y_mean = np.mean(y_pred, axis=0)
y_pred_ind = np.argmax(y_mean)
real_class = os.path.dirname(wav_fn).split("/")[-1]
# print('Actual class: {}, Predicted class: {}'.format(real_class, classes[y_pred]))
for i in range(y_mean.shape[0]):
ret_dict[CLASSES[i]] = f"{y_mean[i]:.4f}"
# ret_dict[CLASSES[y_pred_ind]] = 1
# results.append(y_mean) #{crime: 1, covid: 0, sports: 0, others:0}
return ret_dict
sound = AudioSegment.from_mp3('./aud.mp3')
sound.export('./audio.wav', format='wav')
audio = './audio.wav'
model = load_model('lstm.h5',
custom_objects={
'STFT': STFT,
'Magnitude': Magnitude,
'ApplyFilterbank': ApplyFilterbank,
'MagnitudeToDecibel': MagnitudeToDecibel
})
classes = ['English', 'Hindi', 'Mandarin']
if audio:
st.info('Analyzing audio file')
rate, wav = downsample_mono(audio, 16000)
mask, env = envelope(wav, rate, threshold=10)
clean_wav = wav[mask]
step = int(16000 * 10)
batch = []
for i in range(0, clean_wav.shape[0], step):
sample = clean_wav[i:i + step]
sample = sample.reshape(-1, 1)
if sample.shape[0] < step:
tmp = np.zeros(shape=(step, 1), dtype=np.float32)
tmp[:sample.shape[0], :] = sample.flatten().reshape(-1, 1)
sample = tmp
batch.append(sample)
X_batch = np.array(batch, dtype=np.float32)
y_pred = model.predict(X_batch)
y_mean = np.mean(y_pred, axis=0)