Spaces:
Build error
Build error
Commit
•
5b333be
1
Parent(s):
290e475
Delete app.py
Browse files
app.py
DELETED
@@ -1,178 +0,0 @@
|
|
1 |
-
import gradio as gr
|
2 |
-
import numpy as np
|
3 |
-
import tensorflow as tf
|
4 |
-
from tensorflow import keras
|
5 |
-
import tensorflow_io as tfio
|
6 |
-
from huggingface_hub import from_pretrained_keras
|
7 |
-
|
8 |
-
|
9 |
-
model = from_pretrained_keras("keras-io/ctc_asr", compile=False)
|
10 |
-
|
11 |
-
characters = [x for x in "abcdefghijklmnopqrstuvwxyz'?! "]
|
12 |
-
# Mapping characters to integers
|
13 |
-
char_to_num = keras.layers.StringLookup(vocabulary=characters, oov_token="")
|
14 |
-
# Mapping integers back to original characters
|
15 |
-
num_to_char = keras.layers.StringLookup(
|
16 |
-
vocabulary=char_to_num.get_vocabulary(), oov_token="", invert=True
|
17 |
-
)
|
18 |
-
|
19 |
-
# An integer scalar Tensor. The window length in samples.
|
20 |
-
frame_length = 256
|
21 |
-
# An integer scalar Tensor. The number of samples to step.
|
22 |
-
frame_step = 160
|
23 |
-
# An integer scalar Tensor. The size of the FFT to apply.
|
24 |
-
# If not provided, uses the smallest power of 2 enclosing frame_length.
|
25 |
-
fft_length = 384
|
26 |
-
|
27 |
-
SAMPLE_RATE = 22050
|
28 |
-
|
29 |
-
|
30 |
-
def decode_batch_predictions(pred):
|
31 |
-
input_len = np.ones(pred.shape[0]) * pred.shape[1]
|
32 |
-
# Use greedy search. For complex tasks, you can use beam search
|
33 |
-
results = keras.backend.ctc_decode(pred, input_length=input_len, greedy=True)[0][0]
|
34 |
-
# Iterate over the results and get back the text
|
35 |
-
output_text = []
|
36 |
-
for result in results:
|
37 |
-
result = tf.strings.reduce_join(num_to_char(result)).numpy().decode("utf-8")
|
38 |
-
output_text.append(result)
|
39 |
-
return output_text
|
40 |
-
|
41 |
-
|
42 |
-
def load_16k_audio_wav(filename):
|
43 |
-
# Read file content
|
44 |
-
file_content = tf.io.read_file(filename)
|
45 |
-
|
46 |
-
# Decode audio wave
|
47 |
-
audio_wav, sample_rate = tf.audio.decode_wav(file_content, desired_channels=1)
|
48 |
-
audio_wav = tf.squeeze(audio_wav, axis=-1)
|
49 |
-
sample_rate = tf.cast(sample_rate, dtype=tf.int64)
|
50 |
-
|
51 |
-
# Resample to 16k
|
52 |
-
audio_wav = tfio.audio.resample(
|
53 |
-
audio_wav, rate_in=sample_rate, rate_out=SAMPLE_RATE
|
54 |
-
)
|
55 |
-
|
56 |
-
return audio_wav
|
57 |
-
|
58 |
-
|
59 |
-
def mic_to_tensor(recorded_audio_file):
|
60 |
-
sample_rate, audio = recorded_audio_file
|
61 |
-
|
62 |
-
audio_wav = tf.constant(audio, dtype=tf.float32)
|
63 |
-
if tf.rank(audio_wav) > 1:
|
64 |
-
audio_wav = tf.reduce_mean(audio_wav, axis=1)
|
65 |
-
audio_wav = tfio.audio.resample(
|
66 |
-
audio_wav, rate_in=sample_rate, rate_out=SAMPLE_RATE
|
67 |
-
)
|
68 |
-
|
69 |
-
audio_wav = tf.divide(audio_wav, tf.reduce_max(tf.abs(audio_wav)))
|
70 |
-
|
71 |
-
return audio_wav
|
72 |
-
|
73 |
-
|
74 |
-
def tensor_to_predictions(audio_tensor):
|
75 |
-
# 3. Change type to float
|
76 |
-
audio_tensor = tf.cast(audio_tensor, tf.float32)
|
77 |
-
|
78 |
-
# 4. Get the spectrogram
|
79 |
-
spectrogram = tf.signal.stft(
|
80 |
-
audio_tensor,
|
81 |
-
frame_length=frame_length,
|
82 |
-
frame_step=frame_step,
|
83 |
-
fft_length=fft_length,
|
84 |
-
)
|
85 |
-
|
86 |
-
# 5. We only need the magnitude, which can be derived by applying tf.abs
|
87 |
-
spectrogram = tf.abs(spectrogram)
|
88 |
-
spectrogram = tf.math.pow(spectrogram, 0.5)
|
89 |
-
|
90 |
-
# 6. normalisation
|
91 |
-
means = tf.math.reduce_mean(spectrogram, 1, keepdims=True)
|
92 |
-
stddevs = tf.math.reduce_std(spectrogram, 1, keepdims=True)
|
93 |
-
spectrogram = (spectrogram - means) / (stddevs + 1e-10)
|
94 |
-
|
95 |
-
spectrogram = tf.expand_dims(spectrogram, axis=0)
|
96 |
-
|
97 |
-
batch_predictions = model.predict(spectrogram)
|
98 |
-
batch_predictions = decode_batch_predictions(batch_predictions)
|
99 |
-
return batch_predictions
|
100 |
-
|
101 |
-
|
102 |
-
def clear_inputs_and_outputs():
|
103 |
-
return [None, None, None]
|
104 |
-
|
105 |
-
|
106 |
-
def predict(recorded_audio_file, uploaded_audio_file):
|
107 |
-
# 1. Read wav file
|
108 |
-
if recorded_audio_file:
|
109 |
-
audio_tensor = mic_to_tensor(recorded_audio_file)
|
110 |
-
else:
|
111 |
-
audio_tensor = load_16k_audio_wav(uploaded_audio_file)
|
112 |
-
|
113 |
-
prediction = tensor_to_predictions(audio_tensor)[0]
|
114 |
-
return prediction
|
115 |
-
|
116 |
-
|
117 |
-
# gr.Interface(
|
118 |
-
# infer,
|
119 |
-
# inputs=gr.Audio(source="microphone", type="filepath"),
|
120 |
-
# outputs=gr.Textbox(lines=5, label="Input Text"),
|
121 |
-
# #title=title,
|
122 |
-
# #description=description,
|
123 |
-
# #article=article,
|
124 |
-
# #examples=examples,
|
125 |
-
# enable_queue=True,
|
126 |
-
# ).launch(debug=True)
|
127 |
-
|
128 |
-
# Main function
|
129 |
-
if __name__ == "__main__":
|
130 |
-
demo = gr.Blocks()
|
131 |
-
|
132 |
-
with demo:
|
133 |
-
gr.Markdown(
|
134 |
-
"""
|
135 |
-
<center><h1>Automatic Speech Recognition using CTC</h1></center> \
|
136 |
-
This space is a demo of Automatic Speech Recognition using Keras trained on LJSpeech dataset.<br> \
|
137 |
-
In this space, you can record your voice or upload a wav file and the model will predict the words spoken in English<br><br>
|
138 |
-
"""
|
139 |
-
)
|
140 |
-
with gr.Row():
|
141 |
-
## Input
|
142 |
-
with gr.Column():
|
143 |
-
mic_input = gr.Audio(source="microphone", label="Record your own voice")
|
144 |
-
upl_input = gr.Audio(
|
145 |
-
source="upload", type="filepath", label="Upload a wav file"
|
146 |
-
)
|
147 |
-
|
148 |
-
with gr.Row():
|
149 |
-
clr_btn = gr.Button(value="Clear", variant="secondary")
|
150 |
-
prd_btn = gr.Button(value="Predict")
|
151 |
-
|
152 |
-
# Outputs
|
153 |
-
with gr.Column():
|
154 |
-
lbl_output = gr.Label(label="Text")
|
155 |
-
|
156 |
-
# Credits
|
157 |
-
with gr.Row():
|
158 |
-
gr.Markdown(
|
159 |
-
"""
|
160 |
-
<h4>Credits</h4>
|
161 |
-
Author: <a href="https://twitter.com/anuragcomm"> Anurag Singh</a>.<br>
|
162 |
-
Based on the following Keras example <a href="https://keras.io/examples/audio/ctc_asr">Automatic Speech Recognition using CTC</a> by <a href="https://rbouadjenek.github.io/">Mohamed Reda Bouadjenek</a> and <a href="https://www.linkedin.com/in/parkerhuynh/">Ngoc Dung Huynh</a><br>
|
163 |
-
Check out the model <a href="https://huggingface.co/keras-io/ctc_asr">here</a>
|
164 |
-
"""
|
165 |
-
)
|
166 |
-
|
167 |
-
clr_btn.click(
|
168 |
-
fn=clear_inputs_and_outputs,
|
169 |
-
inputs=[],
|
170 |
-
outputs=[mic_input, upl_input, lbl_output],
|
171 |
-
)
|
172 |
-
prd_btn.click(
|
173 |
-
fn=predict,
|
174 |
-
inputs=[mic_input, upl_input],
|
175 |
-
outputs=[lbl_output],
|
176 |
-
)
|
177 |
-
|
178 |
-
demo.launch(debug=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|