Datasets:
Tasks:
Automatic Speech Recognition
Languages:
English
Upload predict_model.py
Browse files- predict_model.py +71 -0
predict_model.py
ADDED
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from tensorflow import keras
|
2 |
+
import keras.layers
|
3 |
+
import librosa
|
4 |
+
import numpy as np
|
5 |
+
import tensorflow as tf
|
6 |
+
|
7 |
+
frame_length = 256
|
8 |
+
frame_step = 160
|
9 |
+
fft_length = 384
|
10 |
+
|
11 |
+
|
12 |
+
def CTCLoss(y_true, y_pred):
|
13 |
+
batch_len = tf.cast(tf.shape(y_true)[0], dtype="int64")
|
14 |
+
input_length = tf.cast(tf.shape(y_pred)[1], dtype="int64")
|
15 |
+
label_length = tf.cast(tf.shape(y_true)[1], dtype="int64")
|
16 |
+
|
17 |
+
input_length = input_length * tf.ones(shape=(batch_len, 1), dtype="int64")
|
18 |
+
label_length = label_length * tf.ones(shape=(batch_len, 1), dtype="int64")
|
19 |
+
|
20 |
+
loss = keras.backend.ctc_batch_cost(y_true, y_pred, input_length, label_length)
|
21 |
+
return loss
|
22 |
+
|
23 |
+
|
24 |
+
# Tải mô hình
|
25 |
+
loaded_model = keras.models.load_model(r'D:\MyCode\Python\saved_model\my_model.h5', custom_objects={'CTCLoss': CTCLoss})
|
26 |
+
|
27 |
+
characters = [x for x in "abcdefghijklmnopqrstuvwxyzăâêôơưđ'?! "]
|
28 |
+
char_to_num = keras.layers.StringLookup(vocabulary=characters, oov_token="")
|
29 |
+
num_to_char = keras.layers.StringLookup(vocabulary=char_to_num.get_vocabulary(), oov_token="", invert=True)
|
30 |
+
|
31 |
+
|
32 |
+
def decode_batch_predictions(pred):
|
33 |
+
input_len = np.ones(pred.shape[0]) * pred.shape[1]
|
34 |
+
results = keras.backend.ctc_decode(pred, input_len=input_len, greedy=True)[0][0]
|
35 |
+
output_texts = []
|
36 |
+
for result in results:
|
37 |
+
result = tf.strings.reduce_join(num_to_char(result)).numpy().decode('utf-8')
|
38 |
+
output_texts.append(result)
|
39 |
+
return output_texts
|
40 |
+
|
41 |
+
|
42 |
+
# Hàm để xử lý và dự đoán cho một tệp âm thanh
|
43 |
+
def predict_from_audio(file_name):
|
44 |
+
# Tiền xử lý tệp âm thanh
|
45 |
+
audio, _ = librosa.load(file_name, sr=None) # Đọc tệp âm thanh
|
46 |
+
audio = tf.convert_to_tensor(audio, dtype=tf.float32)
|
47 |
+
|
48 |
+
# Tính toán spectrogram
|
49 |
+
spectrogram = tf.signal.stft(audio, frame_length=frame_length, frame_step=frame_step, fft_length=fft_length)
|
50 |
+
spectrogram = tf.abs(spectrogram)
|
51 |
+
spectrogram = tf.math.pow(spectrogram, 0.5)
|
52 |
+
|
53 |
+
# Chuẩn hóa
|
54 |
+
mean = tf.math.reduce_mean(spectrogram, axis=1, keepdims=True)
|
55 |
+
stddevs = tf.math.reduce_std(spectrogram, axis=1, keepdims=True)
|
56 |
+
spectrogram = (spectrogram - mean) / (stddevs + 1e-10)
|
57 |
+
|
58 |
+
# Thêm chiều cho "channels" và "batch"
|
59 |
+
spectrogram = tf.expand_dims(spectrogram, axis=-1) # Thêm chiều cho kênh
|
60 |
+
spectrogram = tf.expand_dims(spectrogram, axis=0) # Thêm chiều batch
|
61 |
+
|
62 |
+
# Dự đoán
|
63 |
+
predictions = loaded_model.predict(spectrogram)
|
64 |
+
decoded_predictions = decode_batch_predictions(predictions)
|
65 |
+
|
66 |
+
return decoded_predictions
|
67 |
+
|
68 |
+
|
69 |
+
# Dự đoán cho một tệp âm thanh
|
70 |
+
result = predict_from_audio(r'D:\MyCode\Python\dataset\test_audio.wav')
|
71 |
+
print("Dự đoán:", result)
|