Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -12,26 +12,47 @@ import matplotlib.pyplot as plt
|
|
12 |
import plotly.express as px
|
13 |
import soundfile as sf
|
14 |
from scipy.signal import stft
|
|
|
15 |
|
16 |
-
#
|
|
|
|
|
17 |
class AudioCNN(nn.Module):
|
18 |
def __init__(self):
|
19 |
super(AudioCNN, self).__init__()
|
|
|
20 |
self.conv1 = nn.Conv2d(1, 16, kernel_size=3, padding=1)
|
21 |
self.conv2 = nn.Conv2d(16, 32, kernel_size=3, padding=1)
|
22 |
-
self.
|
23 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
|
25 |
def forward(self, x):
|
26 |
-
x1 = F.relu(self.conv1(x))
|
27 |
-
x2 =
|
28 |
-
x3 = F.
|
29 |
-
x4 =
|
30 |
-
x5 = F.relu(self.
|
31 |
-
x6 = self.
|
32 |
-
|
33 |
-
|
34 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
35 |
def load_audio(file):
|
36 |
audio, sr = librosa.load(file, sr=None, mono=True)
|
37 |
return audio, sr
|
@@ -53,11 +74,13 @@ def filter_fft(fft, percentage):
|
|
53 |
def create_spectrogram(audio, sr):
|
54 |
n_fft = 2048
|
55 |
hop_length = 512
|
56 |
-
|
57 |
-
spectrogram = np.abs(
|
58 |
return spectrogram, n_fft, hop_length
|
59 |
|
60 |
-
#
|
|
|
|
|
61 |
def plot_waveform(audio, sr, title):
|
62 |
fig = go.Figure()
|
63 |
time = np.arange(len(audio)) / sr
|
@@ -65,41 +88,111 @@ def plot_waveform(audio, sr, title):
|
|
65 |
fig.update_layout(title=title, xaxis_title='Time (s)', yaxis_title='Amplitude')
|
66 |
return fig
|
67 |
|
|
|
|
|
|
|
|
|
|
|
|
|
68 |
def plot_fft(magnitude, phase, sr):
|
69 |
fig = make_subplots(rows=2, cols=1, subplot_titles=('Magnitude Spectrum', 'Phase Spectrum'))
|
70 |
freq = np.fft.fftfreq(len(magnitude), 1/sr)
|
71 |
-
|
72 |
fig.add_trace(go.Scatter(x=freq, y=magnitude, mode='lines', name='Magnitude'), row=1, col=1)
|
73 |
fig.add_trace(go.Scatter(x=freq, y=phase, mode='lines', name='Phase'), row=2, col=1)
|
74 |
-
|
75 |
fig.update_xaxes(title_text='Frequency (Hz)', row=1, col=1)
|
76 |
fig.update_xaxes(title_text='Frequency (Hz)', row=2, col=1)
|
77 |
fig.update_yaxes(title_text='Magnitude', row=1, col=1)
|
78 |
fig.update_yaxes(title_text='Phase (radians)', row=2, col=1)
|
79 |
-
|
80 |
return fig
|
81 |
|
82 |
-
def
|
83 |
freq = np.fft.fftfreq(len(magnitude), 1/sr)
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
88 |
mode='markers',
|
89 |
marker=dict(
|
90 |
-
size=
|
91 |
-
color=
|
92 |
-
colorscale='Viridis',
|
93 |
-
opacity=0.8
|
|
|
94 |
)
|
95 |
-
)
|
96 |
|
|
|
97 |
fig.update_layout(scene=dict(
|
98 |
-
xaxis_title='
|
99 |
-
yaxis_title='
|
100 |
-
zaxis_title='
|
101 |
-
|
102 |
-
|
103 |
return fig
|
104 |
|
105 |
def plot_spectrogram(spectrogram, sr, hop_length):
|
@@ -110,117 +203,185 @@ def plot_spectrogram(spectrogram, sr, hop_length):
|
|
110 |
plt.title('Spectrogram')
|
111 |
return fig
|
112 |
|
113 |
-
def
|
114 |
-
|
115 |
-
df = pd.DataFrame(
|
116 |
-
|
117 |
-
|
118 |
-
'Phase (radians)': phase
|
119 |
-
})
|
120 |
return df
|
121 |
|
122 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
123 |
st.set_page_config(layout="wide")
|
124 |
-
st.title("Audio Frequency Analysis with CNN")
|
125 |
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
|
|
|
|
|
|
|
|
133 |
|
134 |
# File uploader
|
135 |
-
uploaded_file = st.file_uploader("Upload an audio file", type=['wav', 'mp3', 'ogg'])
|
136 |
|
137 |
if uploaded_file is not None:
|
138 |
-
# Load and process audio
|
139 |
audio, sr = load_audio(uploaded_file)
|
140 |
-
st.session_state.audio_data = audio
|
141 |
-
st.session_state.sr = sr
|
142 |
|
143 |
-
#
|
144 |
-
st.
|
145 |
-
st.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
146 |
|
147 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
148 |
fft, magnitude, phase = apply_fft(audio)
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
st.subheader("
|
157 |
-
st.
|
|
|
|
|
158 |
|
159 |
-
#
|
160 |
-
st.
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
|
|
|
|
165 |
percentage = st.slider("Percentage of frequencies to retain:", 0.1, 100.0, 10.0, 0.1)
|
166 |
-
|
167 |
if st.button("Apply Frequency Filter"):
|
168 |
-
filtered_fft = filter_fft(
|
169 |
reconstructed = np.fft.ifft(filtered_fft).real
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
st.audio(reconstructed, sample_rate=sr)
|
177 |
|
178 |
-
# Spectrogram
|
179 |
-
st.
|
|
|
|
|
|
|
|
|
|
|
180 |
spectrogram, n_fft, hop_length = create_spectrogram(audio, sr)
|
181 |
st.pyplot(plot_spectrogram(spectrogram, sr, hop_length))
|
|
|
182 |
|
183 |
-
# CNN
|
184 |
-
|
185 |
-
|
|
|
|
|
|
|
|
|
|
|
186 |
spec_tensor = torch.tensor(spectrogram[np.newaxis, np.newaxis, ...], dtype=torch.float32)
|
187 |
-
|
188 |
model = AudioCNN()
|
189 |
with torch.no_grad():
|
190 |
-
output,
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
200 |
|
201 |
-
|
202 |
-
|
203 |
-
|
|
|
|
|
|
|
|
|
204 |
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
-
plt.tight_layout()
|
213 |
-
st.pyplot(fig)
|
214 |
|
215 |
-
|
216 |
-
|
217 |
-
|
218 |
-
|
219 |
-
|
220 |
-
|
221 |
-
|
222 |
-
|
223 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
224 |
st.markdown("""
|
225 |
<style>
|
226 |
.stButton>button {
|
|
|
12 |
import plotly.express as px
|
13 |
import soundfile as sf
|
14 |
from scipy.signal import stft
|
15 |
+
import math
|
16 |
|
17 |
+
# -------------------------------
|
18 |
+
# CNN Model for Audio Analysis
|
19 |
+
# -------------------------------
|
20 |
class AudioCNN(nn.Module):
|
21 |
def __init__(self):
|
22 |
super(AudioCNN, self).__init__()
|
23 |
+
# Convolutional layers
|
24 |
self.conv1 = nn.Conv2d(1, 16, kernel_size=3, padding=1)
|
25 |
self.conv2 = nn.Conv2d(16, 32, kernel_size=3, padding=1)
|
26 |
+
self.conv3 = nn.Conv2d(32, 64, kernel_size=3, padding=1)
|
27 |
+
# Pooling layer
|
28 |
+
self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
|
29 |
+
# Fully connected layers (with dynamic sizing)
|
30 |
+
self.fc1 = None
|
31 |
+
self.fc2 = nn.Linear(256, 128)
|
32 |
+
self.fc3 = nn.Linear(128, 10)
|
33 |
+
# Dropout for regularization
|
34 |
+
self.dropout = nn.Dropout(0.5)
|
35 |
|
36 |
def forward(self, x):
|
37 |
+
x1 = F.relu(self.conv1(x))
|
38 |
+
x2 = self.pool(x1)
|
39 |
+
x3 = F.relu(self.conv2(x2))
|
40 |
+
x4 = self.pool(x3)
|
41 |
+
x5 = F.relu(self.conv3(x4))
|
42 |
+
x6 = self.pool(x5)
|
43 |
+
if self.fc1 is None:
|
44 |
+
fc1_input_size = x6.numel() // x6.size(0)
|
45 |
+
self.fc1 = nn.Linear(fc1_input_size, 256)
|
46 |
+
x7 = x6.view(x6.size(0), -1)
|
47 |
+
x8 = F.relu(self.fc1(x7))
|
48 |
+
x9 = self.dropout(x8)
|
49 |
+
x10 = F.relu(self.fc2(x9))
|
50 |
+
x11 = self.fc3(x10)
|
51 |
+
return x11, [x2, x4, x6], x8
|
52 |
+
|
53 |
+
# -------------------------------
|
54 |
+
# Audio Processing Functions
|
55 |
+
# -------------------------------
|
56 |
def load_audio(file):
|
57 |
audio, sr = librosa.load(file, sr=None, mono=True)
|
58 |
return audio, sr
|
|
|
74 |
def create_spectrogram(audio, sr):
|
75 |
n_fft = 2048
|
76 |
hop_length = 512
|
77 |
+
S = librosa.stft(audio, n_fft=n_fft, hop_length=hop_length)
|
78 |
+
spectrogram = np.abs(S)
|
79 |
return spectrogram, n_fft, hop_length
|
80 |
|
81 |
+
# -------------------------------
|
82 |
+
# Visualization Functions
|
83 |
+
# -------------------------------
|
84 |
def plot_waveform(audio, sr, title):
|
85 |
fig = go.Figure()
|
86 |
time = np.arange(len(audio)) / sr
|
|
|
88 |
fig.update_layout(title=title, xaxis_title='Time (s)', yaxis_title='Amplitude')
|
89 |
return fig
|
90 |
|
91 |
+
def create_waveform_table(audio, sr, num_samples=100):
|
92 |
+
time = np.arange(len(audio)) / sr
|
93 |
+
indices = np.linspace(0, len(audio)-1, num_samples, dtype=int)
|
94 |
+
df = pd.DataFrame({"Time (s)": time[indices], "Amplitude": audio[indices]})
|
95 |
+
return df
|
96 |
+
|
97 |
def plot_fft(magnitude, phase, sr):
|
98 |
fig = make_subplots(rows=2, cols=1, subplot_titles=('Magnitude Spectrum', 'Phase Spectrum'))
|
99 |
freq = np.fft.fftfreq(len(magnitude), 1/sr)
|
|
|
100 |
fig.add_trace(go.Scatter(x=freq, y=magnitude, mode='lines', name='Magnitude'), row=1, col=1)
|
101 |
fig.add_trace(go.Scatter(x=freq, y=phase, mode='lines', name='Phase'), row=2, col=1)
|
|
|
102 |
fig.update_xaxes(title_text='Frequency (Hz)', row=1, col=1)
|
103 |
fig.update_xaxes(title_text='Frequency (Hz)', row=2, col=1)
|
104 |
fig.update_yaxes(title_text='Magnitude', row=1, col=1)
|
105 |
fig.update_yaxes(title_text='Phase (radians)', row=2, col=1)
|
|
|
106 |
return fig
|
107 |
|
108 |
+
def plot_fft_bands(magnitude, phase, sr):
|
109 |
freq = np.fft.fftfreq(len(magnitude), 1/sr)
|
110 |
+
pos_mask = freq >= 0
|
111 |
+
freq, magnitude, phase = freq[pos_mask], magnitude[pos_mask], phase[pos_mask]
|
112 |
+
bass_mask = (freq >= 20) & (freq < 250)
|
113 |
+
mid_mask = (freq >= 250) & (freq < 4000)
|
114 |
+
treble_mask = (freq >= 4000) & (freq <= sr/2)
|
115 |
+
fig = make_subplots(rows=2, cols=1, subplot_titles=('Magnitude Spectrum by Bands', 'Phase Spectrum by Bands'))
|
116 |
+
fig.add_trace(go.Scatter(x=freq[bass_mask], y=magnitude[bass_mask], mode='lines', name='Bass'), row=1, col=1)
|
117 |
+
fig.add_trace(go.Scatter(x=freq[mid_mask], y=magnitude[mid_mask], mode='lines', name='Mid'), row=1, col=1)
|
118 |
+
fig.add_trace(go.Scatter(x=freq[treble_mask], y=magnitude[treble_mask], mode='lines', name='Treble'), row=1, col=1)
|
119 |
+
fig.add_trace(go.Scatter(x=freq[bass_mask], y=phase[bass_mask], mode='lines', name='Bass'), row=2, col=1)
|
120 |
+
fig.add_trace(go.Scatter(x=freq[mid_mask], y=phase[mid_mask], mode='lines', name='Mid'), row=2, col=1)
|
121 |
+
fig.add_trace(go.Scatter(x=freq[treble_mask], y=phase[treble_mask], mode='lines', name='Treble'), row=2, col=1)
|
122 |
+
fig.update_xaxes(title_text='Frequency (Hz)', row=1, col=1)
|
123 |
+
fig.update_xaxes(title_text='Frequency (Hz)', row=2, col=1)
|
124 |
+
fig.update_yaxes(title_text='Magnitude', row=1, col=1)
|
125 |
+
fig.update_yaxes(title_text='Phase (radians)', row=2, col=1)
|
126 |
+
return fig
|
127 |
+
|
128 |
+
def create_fft_table(magnitude, phase, sr, num_samples=100):
|
129 |
+
freq = np.fft.fftfreq(len(magnitude), 1/sr)
|
130 |
+
pos_mask = freq >= 0
|
131 |
+
freq, magnitude, phase = freq[pos_mask], magnitude[pos_mask], phase[pos_mask]
|
132 |
+
indices = np.linspace(0, len(freq)-1, num_samples, dtype=int)
|
133 |
+
df = pd.DataFrame({
|
134 |
+
"Frequency (Hz)": freq[indices],
|
135 |
+
"Magnitude": magnitude[indices],
|
136 |
+
"Phase (radians)": phase[indices]
|
137 |
+
})
|
138 |
+
return df
|
139 |
+
|
140 |
+
def plot_3d_polar_fft(magnitude, phase, sr):
|
141 |
+
# Get positive frequencies
|
142 |
+
freq = np.fft.fftfreq(len(magnitude), 1/sr)
|
143 |
+
pos_mask = freq >= 0
|
144 |
+
freq, mag, ph = freq[pos_mask], magnitude[pos_mask], phase[pos_mask]
|
145 |
+
# Convert polar to Cartesian coordinates
|
146 |
+
x = mag * np.cos(ph)
|
147 |
+
y = mag * np.sin(ph)
|
148 |
+
z = freq # Use frequency as z-axis
|
149 |
+
|
150 |
+
# Downsample the data to avoid huge message sizes.
|
151 |
+
# Compute a decimation factor so that approximately 500 points are plotted.
|
152 |
+
step = max(1, len(x) // 500)
|
153 |
+
x, y, z, ph = x[::step], y[::step], z[::step], ph[::step]
|
154 |
+
|
155 |
+
# Create a coarser grid for the contour surface.
|
156 |
+
n_rep = 10
|
157 |
+
X_surface = np.tile(x, (n_rep, 1))
|
158 |
+
Y_surface = np.tile(y, (n_rep, 1))
|
159 |
+
Z_surface = np.tile(z, (n_rep, 1))
|
160 |
+
|
161 |
+
surface = go.Surface(
|
162 |
+
x=X_surface,
|
163 |
+
y=Y_surface,
|
164 |
+
z=Z_surface,
|
165 |
+
colorscale='Viridis',
|
166 |
+
opacity=0.6,
|
167 |
+
showscale=False,
|
168 |
+
contours={
|
169 |
+
"x": {"show": True, "start": float(np.min(x)), "end": float(np.max(x)), "size": float((np.max(x)-np.min(x))/10)},
|
170 |
+
"y": {"show": True, "start": float(np.min(y)), "end": float(np.max(y)), "size": float((np.max(y)-np.min(y))/10)},
|
171 |
+
"z": {"show": True, "start": float(np.min(z)), "end": float(np.max(z)), "size": float((np.max(z)-np.min(z))/10)},
|
172 |
+
},
|
173 |
+
)
|
174 |
+
|
175 |
+
scatter = go.Scatter3d(
|
176 |
+
x=x,
|
177 |
+
y=y,
|
178 |
+
z=z,
|
179 |
mode='markers',
|
180 |
marker=dict(
|
181 |
+
size=3,
|
182 |
+
color=ph, # color by phase
|
183 |
+
colorscale='Viridis',
|
184 |
+
opacity=0.8,
|
185 |
+
colorbar=dict(title='Phase (radians)')
|
186 |
)
|
187 |
+
)
|
188 |
|
189 |
+
fig = go.Figure(data=[surface, scatter])
|
190 |
fig.update_layout(scene=dict(
|
191 |
+
xaxis_title='Real Component',
|
192 |
+
yaxis_title='Imaginary Component',
|
193 |
+
zaxis_title='Frequency (Hz)',
|
194 |
+
camera=dict(eye=dict(x=1.5, y=1.5, z=0.5))
|
195 |
+
), margin=dict(l=0, r=0, b=0, t=0))
|
196 |
return fig
|
197 |
|
198 |
def plot_spectrogram(spectrogram, sr, hop_length):
|
|
|
203 |
plt.title('Spectrogram')
|
204 |
return fig
|
205 |
|
206 |
+
def create_spectrogram_table(spectrogram, num_rows=10, num_cols=10):
|
207 |
+
sub_spec = spectrogram[:num_rows, :num_cols]
|
208 |
+
df = pd.DataFrame(sub_spec,
|
209 |
+
index=[f'Freq Bin {i}' for i in range(sub_spec.shape[0])],
|
210 |
+
columns=[f'Time Bin {j}' for j in range(sub_spec.shape[1])])
|
|
|
|
|
211 |
return df
|
212 |
|
213 |
+
def create_activation_table(activation, num_rows=10, num_cols=10):
|
214 |
+
sub_act = activation[:num_rows, :num_cols]
|
215 |
+
df = pd.DataFrame(sub_act,
|
216 |
+
index=[f'Row {i}' for i in range(sub_act.shape[0])],
|
217 |
+
columns=[f'Col {j}' for j in range(sub_act.shape[1])])
|
218 |
+
return df
|
219 |
+
|
220 |
+
# -------------------------------
|
221 |
+
# Streamlit UI & Main App
|
222 |
+
# -------------------------------
|
223 |
st.set_page_config(layout="wide")
|
224 |
+
st.title("Audio Frequency Analysis with CNN and FFT")
|
225 |
|
226 |
+
st.markdown("""
|
227 |
+
### Welcome to the Audio Frequency Analysis Tool!
|
228 |
+
This application allows you to:
|
229 |
+
- **Upload an audio file** and visualize its waveform along with a data table.
|
230 |
+
- **Analyze frequency components** using FFT (with both 2D and enhanced 3D polar plots).
|
231 |
+
- **Highlight frequency bands:** Bass (20–250 Hz), Mid (250–4000 Hz), Treble (4000 Hz to Nyquist).
|
232 |
+
- **Filter frequency components** and reconstruct the waveform.
|
233 |
+
- **Generate a spectrogram** for time-frequency analysis with a sample data table.
|
234 |
+
- **Inspect CNN activations** (pooling and dense layers) arranged in grid layouts.
|
235 |
+
- **Final Audio Classification:** Classify the audio for gender (Male/Female) and tone.
|
236 |
+
""")
|
237 |
|
238 |
# File uploader
|
239 |
+
uploaded_file = st.file_uploader("Upload an audio file (WAV, MP3, OGG)", type=['wav', 'mp3', 'ogg'])
|
240 |
|
241 |
if uploaded_file is not None:
|
|
|
242 |
audio, sr = load_audio(uploaded_file)
|
|
|
|
|
243 |
|
244 |
+
# --- Section 1: Raw Audio Waveform ---
|
245 |
+
st.header("1. Raw Audio Waveform")
|
246 |
+
st.markdown("""
|
247 |
+
The waveform represents the amplitude over time.
|
248 |
+
**Graph:** Amplitude vs. Time.
|
249 |
+
**Data Table:** Sampled values.
|
250 |
+
""")
|
251 |
+
waveform_fig = plot_waveform(audio, sr, "Original Waveform")
|
252 |
+
st.plotly_chart(waveform_fig, use_container_width=True)
|
253 |
+
st.dataframe(create_waveform_table(audio, sr))
|
254 |
|
255 |
+
# --- Section 2: Frequency Domain Analysis ---
|
256 |
+
st.header("2. Frequency Domain Analysis")
|
257 |
+
st.markdown("""
|
258 |
+
**FFT Analysis:** Decompose the audio into frequency components.
|
259 |
+
- **Magnitude Spectrum:** Strength of frequencies.
|
260 |
+
- **Phase Spectrum:** Phase angles.
|
261 |
+
""")
|
262 |
fft, magnitude, phase = apply_fft(audio)
|
263 |
+
col1, col2 = st.columns(2)
|
264 |
+
with col1:
|
265 |
+
st.subheader("2D FFT Plot")
|
266 |
+
st.plotly_chart(plot_fft(magnitude, phase, sr), use_container_width=True)
|
267 |
+
with col2:
|
268 |
+
st.subheader("Enhanced 3D Polar FFT Plot with Contours")
|
269 |
+
st.plotly_chart(plot_3d_polar_fft(magnitude, phase, sr), use_container_width=True)
|
270 |
+
st.subheader("FFT Data Table (Sampled)")
|
271 |
+
st.dataframe(create_fft_table(magnitude, phase, sr))
|
272 |
+
st.subheader("Frequency Bands: Bass, Mid, Treble")
|
273 |
+
st.plotly_chart(plot_fft_bands(magnitude, phase, sr), use_container_width=True)
|
274 |
|
275 |
+
# --- Section 3: Frequency Filtering ---
|
276 |
+
st.header("3. Frequency Filtering")
|
277 |
+
st.markdown("""
|
278 |
+
Filter the audio signal by retaining a percentage of the strongest frequencies.
|
279 |
+
Adjust the slider for retention percentage.
|
280 |
+
**Graph:** Filtered waveform.
|
281 |
+
**Data Table:** Sampled values.
|
282 |
+
""")
|
283 |
percentage = st.slider("Percentage of frequencies to retain:", 0.1, 100.0, 10.0, 0.1)
|
|
|
284 |
if st.button("Apply Frequency Filter"):
|
285 |
+
filtered_fft = filter_fft(fft, percentage)
|
286 |
reconstructed = np.fft.ifft(filtered_fft).real
|
287 |
+
col1, col2 = st.columns(2)
|
288 |
+
with col1:
|
289 |
+
st.plotly_chart(plot_waveform(reconstructed, sr, "Filtered Waveform"), use_container_width=True)
|
290 |
+
with col2:
|
291 |
+
st.audio(reconstructed, sample_rate=sr)
|
292 |
+
st.dataframe(create_waveform_table(reconstructed, sr))
|
|
|
293 |
|
294 |
+
# --- Section 4: Spectrogram Analysis ---
|
295 |
+
st.header("4. Spectrogram Analysis")
|
296 |
+
st.markdown("""
|
297 |
+
A spectrogram shows how frequency content evolves over time.
|
298 |
+
**Graph:** Spectrogram (log-frequency scale).
|
299 |
+
**Data Table:** A subsection of the spectrogram matrix.
|
300 |
+
""")
|
301 |
spectrogram, n_fft, hop_length = create_spectrogram(audio, sr)
|
302 |
st.pyplot(plot_spectrogram(spectrogram, sr, hop_length))
|
303 |
+
st.dataframe(create_spectrogram_table(spectrogram))
|
304 |
|
305 |
+
# --- Section 5: CNN Analysis (Pooling & Dense Activations) ---
|
306 |
+
st.header("5. CNN Analysis: Pooling and Dense Activations")
|
307 |
+
st.markdown("""
|
308 |
+
Instead of classification probabilities, inspect internal activations:
|
309 |
+
- **Pooling Layer Outputs:** Arranged in a grid layout.
|
310 |
+
- **Dense Layer Activation:** Feature vector from the dense layer.
|
311 |
+
""")
|
312 |
+
if st.button("Run CNN Analysis"):
|
313 |
spec_tensor = torch.tensor(spectrogram[np.newaxis, np.newaxis, ...], dtype=torch.float32)
|
|
|
314 |
model = AudioCNN()
|
315 |
with torch.no_grad():
|
316 |
+
output, pooling_outputs, dense_activation = model(spec_tensor)
|
317 |
+
for idx, activation in enumerate(pooling_outputs):
|
318 |
+
st.subheader(f"Pooling Layer {idx+1} Output")
|
319 |
+
act = activation[0].cpu().numpy()
|
320 |
+
num_channels = act.shape[0]
|
321 |
+
ncols = 4
|
322 |
+
nrows = math.ceil(num_channels / ncols)
|
323 |
+
fig, axes = plt.subplots(nrows, ncols, figsize=(3*ncols, 3*nrows))
|
324 |
+
axes = axes.flatten()
|
325 |
+
for i in range(nrows * ncols):
|
326 |
+
if i < num_channels:
|
327 |
+
axes[i].imshow(act[i], aspect='auto', origin='lower', cmap='viridis')
|
328 |
+
axes[i].set_title(f'Channel {i+1}', fontsize=8)
|
329 |
+
axes[i].axis('off')
|
330 |
+
else:
|
331 |
+
axes[i].axis('off')
|
332 |
+
st.pyplot(fig)
|
333 |
+
st.markdown("**Data Table for Pooling Layer Activation (Channel 1, Sampled)**")
|
334 |
+
df_act = create_activation_table(act[0])
|
335 |
+
st.dataframe(df_act)
|
336 |
+
st.subheader("Dense Layer Activation")
|
337 |
+
dense_act = dense_activation[0].cpu().numpy()
|
338 |
+
df_dense = pd.DataFrame({
|
339 |
+
"Feature Index": np.arange(len(dense_act)),
|
340 |
+
"Activation Value": dense_act
|
341 |
+
})
|
342 |
+
st.plotly_chart(px.bar(df_dense, x="Feature Index", y="Activation Value"), use_container_width=True)
|
343 |
+
st.dataframe(df_dense)
|
344 |
+
|
345 |
+
# --- Section 6: Final Audio Classification (Gender & Tone) ---
|
346 |
+
st.header("6. Final Audio Classification: Gender and Tone")
|
347 |
+
st.markdown("""
|
348 |
+
In this final step, a pretrained model classifies the audio as Male or Female,
|
349 |
+
and determines its tone (High Tone vs. Low Tone).
|
350 |
|
351 |
+
**Note:** This example uses a placeholder model. Replace the dummy model and random outputs with your actual pretrained model.
|
352 |
+
""")
|
353 |
+
if st.button("Run Final Classification"):
|
354 |
+
# Extract MFCC features as an example (adjust as needed)
|
355 |
+
mfccs = librosa.feature.mfcc(y=audio, sr=sr, n_mfcc=40)
|
356 |
+
features = np.mean(mfccs, axis=1) # average over time
|
357 |
+
features_tensor = torch.tensor(features, dtype=torch.float32).unsqueeze(0)
|
358 |
|
359 |
+
# Dummy classifier model for demonstration
|
360 |
+
class GenderToneClassifier(nn.Module):
|
361 |
+
def __init__(self):
|
362 |
+
super(GenderToneClassifier, self).__init__()
|
363 |
+
self.fc = nn.Linear(40, 4) # 4 outputs: [Male, Female, High Tone, Low Tone]
|
364 |
+
def forward(self, x):
|
365 |
+
return self.fc(x)
|
|
|
|
|
366 |
|
367 |
+
classifier = GenderToneClassifier()
|
368 |
+
# In practice, load your pretrained weights here.
|
369 |
+
with torch.no_grad():
|
370 |
+
output = classifier(features_tensor)
|
371 |
+
probs = F.softmax(output, dim=1).numpy()[0]
|
372 |
+
# Interpret outputs: assume first 2 are gender, next 2 are tone.
|
373 |
+
gender = "Male" if probs[0] > probs[1] else "Female"
|
374 |
+
tone = "High Tone" if probs[2] > probs[3] else "Low Tone"
|
375 |
+
st.markdown(f"**Predicted Gender:** {gender}")
|
376 |
+
st.markdown(f"**Predicted Tone:** {tone}")
|
377 |
+
categories = ["Male", "Female", "High Tone", "Low Tone"]
|
378 |
+
df_class = pd.DataFrame({"Category": categories, "Probability": probs})
|
379 |
+
st.plotly_chart(px.bar(df_class, x="Category", y="Probability"), use_container_width=True)
|
380 |
+
st.dataframe(df_class)
|
381 |
+
|
382 |
+
# -------------------------------
|
383 |
+
# Style Enhancements
|
384 |
+
# -------------------------------
|
385 |
st.markdown("""
|
386 |
<style>
|
387 |
.stButton>button {
|