Update app.py

app.py

@@ -19,9 +19,19 @@ def extract_mel_spectrogram(audio) -> np.ndarray:
     #spectrogram = np.expand_dims(spectrogram, axis=-1)  # Adding channel dimension for the model
     return spectrogram
 
+def CTCLoss(y_true, y_pred):
+    # Compute the training-time loss value
+    batch_len = tf.cast(tf.shape(y_true)[0], dtype="int64")
+    input_length = tf.math.reduce_sum(tf.cast(tf.not_equal(tf.reduce_max(y_pred, axis=2), 0), dtype="int64"), axis=1, keepdims=True)
+    label_length = tf.math.reduce_sum(tf.cast(tf.not_equal(y_true, -1), dtype="int64"), axis=1, keepdims=True)
+
+    loss = tf.keras.backend.ctc_batch_cost(y_true, y_pred, input_length, label_length)
+    return loss
+
 # Download model from Hugging Face Hub
 model_path = hf_hub_download(repo_id="kobrasoft/kobraspeech-rnn-cs", filename="kobraspeech.17-40.19.keras")
-model = tf.keras.models.load_model(model_path)
+with tf.keras.utils.custom_object_scope({'CTCLoss': CTCLoss}):
+    model = tf.keras.models.load_model(model_path)
 
 def decode_batch_predictions(pred):
     input_len = np.ones(pred.shape[0]) * pred.shape[1]