Here is my architecture.

def model1():

    inputs = layers.Input(shape=(x1.shape[-2],x1.shape[-1]), name='Input_layer')

    x = layers.Reshape((x1.shape[-2], x1.shape[-1], 1), name='Input_reshape')(inputs)
    #x= layers.Masking(mask_value=0.0, input_shape=(109, 80))(x)

    x = layers.Conv2D(32, 3, strides=2,
                  activation=layers.LeakyReLU(alpha=0.2), 
                  name='convolution_1')(x)
    x = layers.BatchNormalization()(x)
    x = layers.Dropout(0.2, name='dropout1_model1')(x)

    x = layers.Conv2D(64, 3, strides=1,
                  activation=layers.LeakyReLU(alpha=0.2),
                  name='convolution_2')(x)
    x = layers.BatchNormalization()(x)
    x = layers.Dropout(0.2, name='dropout2_model1')(x)

    x = layers.Reshape((-1, x.shape[-2] * x.shape[-1]), name='conv_reshpae')(x)

    x = layers.Bidirectional(layers.GRU(64, return_sequences=True, activation=layers.LeakyReLU(alpha=0.2), name='GRU_1'))(x)
    x = layers.Dropout(0.2, name='dropout3_model1')(x)

    x = layers.Bidirectional(layers.GRU(64, return_sequences=True, activation=layers.LeakyReLU(alpha=0.2), name='GRU_2'))(x)
    x = layers.Dropout(0.2, name='dropout4_model1')(x)

    outputs=layers.Dense(128, activation=layers.LeakyReLU(alpha=0.2), name='output_1')(x)

    model=keras.Model(inputs=inputs, outputs=outputs, name='audio_model')

    return model

model_audio=model1()
print(model_audio.summary())

def model2():

  inputs=layers.Input(shape=x2.shape[-1])
  x=layers.Embedding(vocab_size, embedding_dim, mask_zero=True)(inputs)
  x=layers.Bidirectional(LSTM(64, return_sequences=True, activation=layers.LeakyReLU(alpha=0.2), name='First_LSTM'))(x)
  x = layers.Dropout(0.2, name='dropout1_model2')(x)
  x=layers.Bidirectional(LSTM(64, return_sequences=True, activation=layers.LeakyReLU(alpha=0.2), name='Second_LSTM'))(x)
  x = layers.Dropout(0.2, name='dropout2_model')(x)
  outputs=layers.Dense(128, activation=layers.LeakyReLU(alpha=0.2), name='output_2')(x)
  model=keras.Model(inputs=inputs,outputs=outputs, name='text_model')
  return model

model_text=model2()
print(model_text.summary())

def model3(model_audio,model_text):

  Input_1 = tf.keras.layers.Input(shape=(x1.shape[-2], x1.shape[-1]), name='Input_1')
  Input_2 = tf.keras.layers.Input(shape=(x2.shape[-1]), name='Input_2')

  output_1 = model_audio(Input_1)
  output_2 = model_text(Input_2)

  layer = layers.MultiHeadAttention(num_heads=2, key_dim=1, name='attention_layer')
  context_vector, attention_weights = layer(output_1, output_2, return_attention_scores=True)
  x = layers.GRU(128, return_sequences=False, activation=layers.LeakyReLU(alpha=0.2), name='GRU')(context_vector)
  x = layers.Dropout(0.2, name='dropout1_modelFinal')(x)
  x = layers.Dense(32, activation=layers.LeakyReLU(alpha=0.2))(x)
  x = layers.Dropout(0.2, name='dropout2_modelFinal')(x)
  outputs = layers.Dense(1,activation='sigmoid', name='final_layer')(x)

  model = keras.Model(inputs=[Input_1, Input_2], outputs=outputs)
  return model

model_final = model3(model_audio,model_text)
print(model_final.summary())

Here is the model3 (model final) summary

Layer (type) Output Shape Param # Connected to

Input_1 (InputLayer) [(None, 135, 40)] 0 []

Input_2 (InputLayer) [(None, 17)] 0 []

model (Functional) (None, 65, 128) 553344 ['Input_1[0][0]']

model_1 (Functional) (None, 17, 128) 287616 ['Input_2[0][0]']

attention_layer (MultiHead ((None, 65, 128), 1158 ['model[0][0]',
Attention) (None, 2, 65, 17)) 'model_1[0][0]']

GRU (GRU) (None, 128) 99072 ['attention_layer[0][0]']

dropout1_modelFinal (Dropo (None, 128) 0 ['GRU[0][0]']
ut)

dense (Dense) (None, 32) 4128 ['dropout1_modelFinal[0][0]']

dropout2_modelFinal (Dropo (None, 32) 0 ['dense[0][0]']
ut)

final_layer (Dense) (None, 1) 33 ['dropout2_modelFinal[0][0]']

================================================================================================== Total params: 945351 (3.61 MB) Trainable params: 945159 (3.61 MB) Non-trainable params: 192 (768.00 Byte)

Here I want to extract the model (Functional) output which is an another model (mode_audio) defined. It is a submodel fused in main model. actually model_audio output is used an input to model_final

Initially this is what tried

model_output = model.layers[-8].output  # 'model' layer is at index -8 from the summary

model_layer_output = Model(inputs=model.input, outputs=model_output)

intermediate_output = model_layer_output.predict([test_input_1_data, test_input_2_data])

since layer[-8] is the model_audio layer .i used that as an output. But i was getting the below error

Graph disconnected: cannot obtain value for tensor KerasTensor(type_spec=TensorSpec(shape=(None, 135, 40), dtype=tf.float32, name='Input_layer'), name='Input_layer', description="created by layer 'Input_layer'") at layer "Input_reshape". The following previous layers were accessed without issue: []

So later i tried this. I got the expected output shape. but i am not sure whether it is right

#audio model
input_layer = Input(shape=(135, 40))  
model_audio_output = model.get_layer('model')(input_layer)

submodel = Model(inputs=input_layer, outputs=model_audio_output)
audio_embed=submodel.predict(input_samples)

can someone suggest how to extract the model_audio output from my saved keras model