def _get_output_tensors(
self, interpreter: tf.lite.Interpreter) -> List[np.ndarray]:
"""Returns output tensors of given TFLite model Interpreter.
Args:
interpreter: a tf.lite.Interpreter object with allocated tensors.
Returns:
a list of numpy arrays representing output tensor results.
"""
return [
interpreter.get_tensor(tensor['index'])
for tensor in interpreter.get_output_details()
]
def _get_output_tensors(
self, interpreter: tf.lite.Interpreter) -> List[np.ndarray]:
"""Returns output tensors of given TFLite model Interpreter.
Args:
interpreter: a tf.lite.Interpreter object with allocated tensors.
Returns:
a list of numpy arrays representing output tensor results.
"""
outputs = []
for output_detail in interpreter.get_output_details():
tensor = interpreter.get_tensor(output_detail['index'])
if output_detail['dtype'] == np.int8:
quant_params = _get_quant_params(output_detail)
if quant_params:
scale, zero_point = quant_params
tensor = ((tensor.astype(np.float32) - zero_point) *
scale).astype(np.float32)
outputs.append(tensor)
return outputs
def clean_speech(audio, interpreter_1: tf.lite.Interpreter,
interpreter_2: tf.lite.Interpreter):
block_len = 512
block_shift = 128
# load models
interpreter_1.allocate_tensors()
interpreter_2.allocate_tensors()
# Get input and output tensors.
input_details_1 = interpreter_1.get_input_details()
output_details_1 = interpreter_1.get_output_details()
input_details_2 = interpreter_2.get_input_details()
output_details_2 = interpreter_2.get_output_details()
# create states for the lstms
states_1 = np.zeros(input_details_1[1]['shape']).astype('float32')
states_2 = np.zeros(input_details_2[1]['shape']).astype('float32')
# preallocate output audio
out_file = np.zeros((len(audio)))
# create buffer
in_buffer = np.zeros((block_len)).astype('float32')
out_buffer = np.zeros((block_len)).astype('float32')
# calculate number of blocks
num_blocks = (audio.shape[0] - (block_len - block_shift)) // block_shift
# iterate over the number of blcoks
for idx in range(num_blocks):
# shift values and write to buffer
in_buffer[:-block_shift] = in_buffer[block_shift:]
in_buffer[-block_shift:] = audio[idx *
block_shift:(idx * block_shift) +
block_shift]
# calculate fft of input block
in_block_fft = np.fft.rfft(in_buffer)
in_mag = np.abs(in_block_fft)
in_phase = np.angle(in_block_fft)
# reshape magnitude to input dimensions
in_mag = np.reshape(in_mag, (1, 1, -1)).astype('float32')
# set tensors to the first model
interpreter_1.set_tensor(input_details_1[1]['index'], states_1)
interpreter_1.set_tensor(input_details_1[0]['index'], in_mag)
# run calculation
interpreter_1.invoke()
# get the output of the first block
out_mask = interpreter_1.get_tensor(output_details_1[0]['index'])
states_1 = interpreter_1.get_tensor(output_details_1[1]['index'])
# calculate the ifft
estimated_complex = in_mag * out_mask * np.exp(1j * in_phase)
estimated_block = np.fft.irfft(estimated_complex)
# reshape the time domain block
estimated_block = np.reshape(estimated_block,
(1, 1, -1)).astype('float32')
# set tensors to the second block
interpreter_2.set_tensor(input_details_2[1]['index'], states_2)
interpreter_2.set_tensor(input_details_2[0]['index'], estimated_block)
# run calculation
interpreter_2.invoke()
# get output tensors
out_block = interpreter_2.get_tensor(output_details_2[0]['index'])
states_2 = interpreter_2.get_tensor(output_details_2[1]['index'])
# shift values and write to buffer
out_buffer[:-block_shift] = out_buffer[block_shift:]
out_buffer[-block_shift:] = np.zeros((block_shift))
out_buffer += np.squeeze(out_block)
# write block to output file
out_file[idx * block_shift:(idx * block_shift) +
block_shift] = out_buffer[:block_shift]
output_bytes = io.BytesIO()
sf.write('out.wav', out_file, 16000)
return output_bytes