def infer(outputs, i):
predicted_final_specs = outputs[1]
sequence_lengths = outputs[4]
for j in range(len(predicted_final_specs)):
predicted_final_spec = predicted_final_specs[j]
audio_length = sequence_lengths[j]
if audio_length > 2:
if "both" in model.get_data_layer().params['output_type']:
predicted_mag_spec = outputs[5][j][:audio_length -
1, :]
else:
predicted_final_spec = predicted_final_spec[:
audio_length
- 1, :]
predicted_mag_spec = model.get_data_layer(
).get_magnitude_spec(predicted_final_spec, is_mel=True)
save_audio(
predicted_mag_spec,
syn_save_dir,
0,
n_fft=model.get_data_layer().n_fft,
sampling_rate=model.get_data_layer().sampling_rate,
mode="syn",
number=i * batch_size + j,
save_format="disk",
gl_iters=4,
verbose=False)
else:
print(
"WARNING: An audio file was not saved, this will error out in"
"future steps")
def infer(line):
print("Input English")
print(line)
# Generate speech
results = get_interactive_infer_results(model_T2S, sess, model_in=[line])
audio_length = results[1][4][0]
if model_T2S.get_data_layer()._both:
prediction = results[1][5][0]
else:
prediction = results[1][1][0]
prediction = prediction[:audio_length-1,:]
mag_prediction = model_T2S.get_data_layer().get_magnitude_spec(prediction)
mag_prediction_squared = np.clip(mag_prediction, a_min=0, a_max=255)
mag_prediction_squared = mag_prediction_squared**1.5
mag_prediction_squared = np.square(mag_prediction_squared)
mel_basis = librosa.filters.mel(sr=22050, n_fft=1024, n_mels=80, htk=True, norm=None)
mel = np.dot(mel_basis, mag_prediction_squared.T)
mel = np.log(np.clip(mel, a_min=1e-5, a_max=None))
np.save("spec2", mel)
plt.imshow(mel)
plt.gca().invert_yaxis()
plt.show()
wav = save_audio(mag_prediction, "unused", "unused", sampling_rate=sampling_rate, save_format="np.array", n_fft=n_fft)
audio = IPython.display.Audio(wav, rate=sampling_rate)
print("Generated Audio")
IPython.display.display(audio)
def infer(line):
print("Input English")
print(line)
# Generate speech
results = get_t2s_interactive_infer_results(model_T2S, model_in=[line])
audio_length = results[1][4][0]
if model_T2S.get_data_layer()._both:
prediction = results[1][5][0]
else:
prediction = results[1][1][0]
prediction = prediction[:audio_length - 1, :]
mag_prediction = model_T2S.get_data_layer().get_magnitude_spec(prediction)
wav = save_audio(mag_prediction,
"unused",
"unused",
sampling_rate=sampling_rate,
save_format="np.array",
n_fft=n_fft)
wav = librosa.core.resample(wav, sampling_rate, 16000)
print("Generated Audio")
# Generate text
results = get_s2t_interactive_infer_results(model_S2T, model_in=[wav])
english_recognized = results[0][0]
print("Recognized Speech")
print(english_recognized)
# Generate translation
encoded_src_list = sp1.EncodeAsPieces(english_recognized)
encoded_src = ' '.join([w for w in encoded_src_list])
results = get_t2t_interactive_infer_results(model_T2T,
model_in=[encoded_src])
encoded_tgt = results[1][0]
decoded_tgt = sp2.DecodePieces(encoded_tgt.split(" "))
print("Translation")
print(decoded_tgt)
def infer(line):
print("Input English")
print(line)
# Generate speech
results = get_interactive_infer_results(model_T2S, model_in=[line])
audio_length = results[1][4][0]
if model_T2S.get_data_layer()._both:
prediction = results[1][5][0]
else:
prediction = results[1][1][0]
prediction = prediction[:audio_length - 1, :]
mag_prediction = model_T2S.get_data_layer().get_magnitude_spec(prediction)
mag_prediction_squared = np.clip(mag_prediction, a_min=0, a_max=255)
mag_prediction_squared = mag_prediction_squared**1.5
mag_prediction_squared = np.square(mag_prediction_squared)
mel_basis = librosa.filters.mel(sr=22050,
n_fft=1024,
n_mels=80,
htk=True,
norm=None)
mel = np.dot(mel_basis, mag_prediction_squared.T)
mel = np.log(np.clip(mel, a_min=1e-5, a_max=None))
np.save("spec2", mel)
wav = save_audio(mag_prediction,
"/home/oscar/filesys/tacotron-LJ-float/checkpoint/",
"1",
sampling_rate=sampling_rate,
save_format="disk",
n_fft=n_fft)
print("Generated Audio")