Datasets:
Tasks:
Automatic Speech Recognition
Languages:
English
Upload speech2text.py
Browse files- speech2text.py +243 -0
speech2text.py
ADDED
@@ -0,0 +1,243 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
import keras.layers
|
4 |
+
import librosa
|
5 |
+
import matplotlib.pyplot as plt
|
6 |
+
import numpy as np
|
7 |
+
import pandas as pd
|
8 |
+
from jiwer import wer
|
9 |
+
from keras.src.applications.densenet import layers
|
10 |
+
from scipy.io import wavfile
|
11 |
+
import tensorflow as tf
|
12 |
+
|
13 |
+
data_path = r"D:\MyCode\Python\dataset\LJSpeech-1.1"
|
14 |
+
wave_path = data_path + "/wavs/"
|
15 |
+
metadata_path = data_path + '/metadata.csv'
|
16 |
+
|
17 |
+
metadata_df = pd.read_csv(metadata_path, sep="|", header=None, quoting=3)
|
18 |
+
metadata_df.columns = ["file_name", "transcription", "normalized_transcription"]
|
19 |
+
metadata_df = metadata_df[["file_name", "transcription", "normalized_transcription"]]
|
20 |
+
metadata_df = metadata_df.sample(frac=1).reset_index(drop=True)
|
21 |
+
print(metadata_df.head(10))
|
22 |
+
|
23 |
+
split = int(len(metadata_df) * 0.90)
|
24 |
+
df_train = metadata_df[:split]
|
25 |
+
df_test = metadata_df[split:]
|
26 |
+
|
27 |
+
frame_length = 256
|
28 |
+
frame_step = 160
|
29 |
+
fft_length = 384
|
30 |
+
|
31 |
+
batch_size = 32
|
32 |
+
epochs = 10
|
33 |
+
|
34 |
+
# preprocessing
|
35 |
+
characters = [x for x in "abcdefghijklmnopqrstuvwxyzăâêôơưđ'?! "]
|
36 |
+
char_to_num = keras.layers.StringLookup(vocabulary=characters, oov_token="")
|
37 |
+
num_to_char = keras.layers.StringLookup(vocabulary=char_to_num.get_vocabulary(), oov_token="", invert=True)
|
38 |
+
|
39 |
+
|
40 |
+
# def encode_single_sample(wav_file, label):
|
41 |
+
# file = tf.io.read_file(wave_path, wav_file + ".wav")
|
42 |
+
# audio, _ = tf.audio.decode_wav(file)
|
43 |
+
# audio = tf.squeeze(audio, axis=-1)
|
44 |
+
# audio = tf.cast(audio, tf.float32)
|
45 |
+
#
|
46 |
+
# spectrogram = tf.signal.stft(audio, frame_length=frame_length, frame_step=frame_step, fft_length=fft_length)
|
47 |
+
# spectrogram = tf.abs(spectrogram)
|
48 |
+
# spectrogram = tf.math.pow(spectrogram, 0.5)
|
49 |
+
#
|
50 |
+
# mean = tf.math.reduce_mean(spectrogram, 1, keepdims=True)
|
51 |
+
# stddevs = tf.math.reduce_std(spectrogram, 1, keepdims=True)
|
52 |
+
# spectrogram = (spectrogram - mean) / (stddevs + 1e-10)
|
53 |
+
#
|
54 |
+
# label = tf.strings.lower(label)
|
55 |
+
# label = tf.strings.unicode_split(label, input_encoding='UTF-8')
|
56 |
+
# label = char_to_num(label)
|
57 |
+
# return spectrogram, label
|
58 |
+
|
59 |
+
|
60 |
+
def encode_single_sample(wav_file, label):
|
61 |
+
# Tạo đường dẫn file âm thanh
|
62 |
+
file_path = tf.strings.join([wave_path, wav_file, ".wav"], separator="")
|
63 |
+
|
64 |
+
# Đọc file âm thanh
|
65 |
+
file = tf.io.read_file(file_path)
|
66 |
+
audio, _ = tf.audio.decode_wav(file)
|
67 |
+
audio = tf.squeeze(audio, axis=-1)
|
68 |
+
audio = tf.cast(audio, tf.float32)
|
69 |
+
|
70 |
+
# Tính toán spectrogram
|
71 |
+
spectrogram = tf.signal.stft(audio, frame_length=frame_length, frame_step=frame_step, fft_length=fft_length)
|
72 |
+
spectrogram = tf.abs(spectrogram)
|
73 |
+
spectrogram = tf.math.pow(spectrogram, 0.5)
|
74 |
+
|
75 |
+
# Chuẩn hóa
|
76 |
+
mean = tf.math.reduce_mean(spectrogram, axis=1, keepdims=True)
|
77 |
+
stddevs = tf.math.reduce_std(spectrogram, axis=1, keepdims=True)
|
78 |
+
spectrogram = (spectrogram - mean) / (stddevs + 1e-10)
|
79 |
+
|
80 |
+
# Thêm chiều cho "channels"
|
81 |
+
spectrogram = tf.expand_dims(spectrogram, axis=-1) # Giữ nguyên
|
82 |
+
spectrogram = tf.expand_dims(spectrogram, axis=0) # Thêm chiều batch
|
83 |
+
|
84 |
+
# Xử lý nhãn
|
85 |
+
label = tf.strings.lower(label)
|
86 |
+
label = tf.strings.unicode_split(label, input_encoding='UTF-8')
|
87 |
+
label = char_to_num(label)
|
88 |
+
return spectrogram, label
|
89 |
+
|
90 |
+
|
91 |
+
train_dataset = tf.data.Dataset.from_tensor_slices((
|
92 |
+
list(df_train["file_name"]),
|
93 |
+
list(df_train["normalized_transcription"])
|
94 |
+
))
|
95 |
+
|
96 |
+
train_dataset = (
|
97 |
+
train_dataset.map(encode_single_sample, num_parallel_calls=tf.data.AUTOTUNE)
|
98 |
+
.padded_batch(batch_size)
|
99 |
+
.prefetch(buffer_size=tf.data.AUTOTUNE)
|
100 |
+
)
|
101 |
+
|
102 |
+
# Tạo dataset cho validation
|
103 |
+
validation_dataset = tf.data.Dataset.from_tensor_slices((
|
104 |
+
list(df_test["file_name"]),
|
105 |
+
list(df_test["normalized_transcription"])
|
106 |
+
))
|
107 |
+
|
108 |
+
validation_dataset = (
|
109 |
+
validation_dataset.map(encode_single_sample, num_parallel_calls=tf.data.AUTOTUNE)
|
110 |
+
.padded_batch(batch_size)
|
111 |
+
.prefetch(buffer_size=tf.data.AUTOTUNE)
|
112 |
+
)
|
113 |
+
|
114 |
+
for batch in train_dataset.take(1):
|
115 |
+
spectrogram = batch[0][0].numpy() # Lấy spectrogram từ batch
|
116 |
+
|
117 |
+
# Kiểm tra kích thước
|
118 |
+
if spectrogram.ndim == 4: # Nếu là mảng 4D, loại bỏ chiều batch
|
119 |
+
spectrogram = tf.squeeze(spectrogram, axis=0)
|
120 |
+
|
121 |
+
# Kiểm tra lại nếu là mảng 3D
|
122 |
+
if spectrogram.ndim == 3: # Nếu vẫn là mảng 3D, chuyển đổi về mảng 2D
|
123 |
+
spectrogram = np.squeeze(spectrogram, axis=-1) # Chuyển đổi về mảng 2D
|
124 |
+
|
125 |
+
# Áp dụng np.trim_zeros cho từng hàng
|
126 |
+
trimmed_spectrogram = [np.trim_zeros(x) for x in spectrogram.T] # Chuyển vị và trim
|
127 |
+
|
128 |
+
# Chuyển đổi về numpy array 2D nếu cần
|
129 |
+
max_length = max(len(x) for x in trimmed_spectrogram) # Tìm chiều dài tối đa
|
130 |
+
trimmed_spectrogram = np.array([np.pad(x, (0, max_length - len(x)), mode='constant') for x in trimmed_spectrogram])
|
131 |
+
|
132 |
+
|
133 |
+
def CTCLoss(y_true, y_pred):
|
134 |
+
batch_len = tf.cast(tf.shape(y_true)[0], dtype="int64")
|
135 |
+
input_length = tf.cast(tf.shape(y_pred)[1], dtype="int64")
|
136 |
+
label_length = tf.cast(tf.shape(y_true)[1], dtype="int64")
|
137 |
+
|
138 |
+
input_length = input_length * tf.ones(shape=(batch_len, 1), dtype="int64")
|
139 |
+
label_length = label_length * tf.ones(shape=(batch_len, 1), dtype="int64")
|
140 |
+
|
141 |
+
loss = keras.backend.ctc_batch_cost(y_true, y_pred, input_length, label_length)
|
142 |
+
return loss
|
143 |
+
|
144 |
+
|
145 |
+
def build_model(input_dim, output_dim, rnn_layer=5, rnn_units=128):
|
146 |
+
input_spectrogram = layers.Input(shape=(None, input_dim), name="input")
|
147 |
+
|
148 |
+
x = layers.Reshape((-1, input_dim, 1), name="expand_dim")(input_spectrogram)
|
149 |
+
|
150 |
+
# Lớp Convolutional 1
|
151 |
+
x = layers.Conv2D(filters=32, kernel_size=[11, 41], strides=[2, 2],
|
152 |
+
padding="same", use_bias=False, name="conv_1")(x)
|
153 |
+
x = layers.BatchNormalization(name="bn_conv_1")(x) # Đổi tên lớp này
|
154 |
+
x = layers.ReLU(name="relu_1")(x)
|
155 |
+
|
156 |
+
# Lớp Convolutional 2
|
157 |
+
x = layers.Conv2D(filters=32, kernel_size=[11, 21], strides=[1, 2],
|
158 |
+
padding="same", use_bias=False, name="conv_2")(x)
|
159 |
+
x = layers.BatchNormalization(name="bn_conv_2")(x) # Đổi tên lớp này
|
160 |
+
|
161 |
+
x = layers.ReLU(name="relu_2")(x)
|
162 |
+
|
163 |
+
# Reshape để sử dụng với RNN
|
164 |
+
x = layers.Reshape((-1, x.shape[-2] * x.shape[-1]))(x)
|
165 |
+
|
166 |
+
for i in range(1, rnn_layer + 1):
|
167 |
+
recurrent = layers.GRU(
|
168 |
+
units=rnn_units,
|
169 |
+
activation="tanh",
|
170 |
+
recurrent_activation="sigmoid",
|
171 |
+
use_bias=True,
|
172 |
+
return_sequences=True,
|
173 |
+
reset_after=True,
|
174 |
+
name=f"gru_{i}",
|
175 |
+
)
|
176 |
+
# Các lớp Recurrent
|
177 |
+
x = layers.Bidirectional(
|
178 |
+
recurrent, name=f"bidirectional_{i}", merge_mode="concat",
|
179 |
+
)(x)
|
180 |
+
if i < rnn_layer:
|
181 |
+
x = layers.Dropout(rate=0.5)(x)
|
182 |
+
|
183 |
+
x = layers.Dense(units=rnn_units * 2, name="dense_1")(x)
|
184 |
+
x = layers.ReLU(name="relu_3")(x)
|
185 |
+
x = layers.Dropout(rate=0.5)(x)
|
186 |
+
|
187 |
+
output = layers.Dense(units=output_dim + 1, activation="softmax")(x)
|
188 |
+
model = keras.Model(input_spectrogram, output, name="DeepSpeech_2")
|
189 |
+
otp = keras.optimizers.Adam(learning_rate=1e-4)
|
190 |
+
model.compile(optimizer=otp, loss=CTCLoss)
|
191 |
+
|
192 |
+
return model
|
193 |
+
|
194 |
+
|
195 |
+
model = build_model(
|
196 |
+
input_dim=fft_length // 2 + 1,
|
197 |
+
output_dim=char_to_num.vocab_size(),
|
198 |
+
rnn_units=512,
|
199 |
+
)
|
200 |
+
model.summary()
|
201 |
+
|
202 |
+
|
203 |
+
def decode_batch_predictions(pred):
|
204 |
+
input_len = np.ones(pred.shape[0]) * pred.shape[1]
|
205 |
+
results = keras.backend.ctc_decode(pred, input_len=input_len, greedy=True)[0][0]
|
206 |
+
output_texts = []
|
207 |
+
for result in results:
|
208 |
+
result = tf.strings.reduce_join(num_to_char(result)).numpy().decode('utf-8')
|
209 |
+
output_texts.append(result)
|
210 |
+
return output_texts
|
211 |
+
|
212 |
+
|
213 |
+
class CallbackEval(keras.callbacks.Callback):
|
214 |
+
def __init__(self, dataset):
|
215 |
+
super().__init__()
|
216 |
+
self.dataset = dataset
|
217 |
+
|
218 |
+
def on_epoch_end(self, epoch, logs=None):
|
219 |
+
prediction = []
|
220 |
+
targets = []
|
221 |
+
for batch in self.dataset:
|
222 |
+
X, y = batch
|
223 |
+
batch_predictions = model.predict(X)
|
224 |
+
batch_predictions = decode_batch_predictions(batch_predictions)
|
225 |
+
prediction.extend(batch_predictions)
|
226 |
+
for label in y:
|
227 |
+
label = (tf.strings.reduce_join(num_to_char(label)).numpy().decode("utf-8"))
|
228 |
+
targets.append(label)
|
229 |
+
wer_score = wer(targets, prediction)
|
230 |
+
print(f"WER: {wer_score:.4f}")
|
231 |
+
for i in np.random.randint(0, len(prediction), 2):
|
232 |
+
print(f"Target: {targets[i]}")
|
233 |
+
print(f"Prediction: {prediction[i]}")
|
234 |
+
|
235 |
+
|
236 |
+
validation_callback = CallbackEval(validation_dataset)
|
237 |
+
history = model.fit(
|
238 |
+
train_dataset,
|
239 |
+
validation_data=validation_dataset,
|
240 |
+
epochs=epochs,
|
241 |
+
callbacks=[validation_callback],
|
242 |
+
)
|
243 |
+
model.save(r'D:\MyCode\Python\pythonProject\SavedModed\model_stt.h5')
|