Spaces:
Running
on
T4
Running
on
T4
Update app.py
Browse files
app.py
CHANGED
@@ -13,12 +13,13 @@ import os
|
|
13 |
DEFAULT_TOKEN_RATE = 100
|
14 |
DEFAULT_SEMANTIC_VOCAB_SIZE = 16384
|
15 |
DEFAULT_SAMPLE_RATE = 16000
|
|
|
16 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
17 |
|
18 |
# Title and Description
|
19 |
st.title("SemantiCodec: Ultra-Low Bitrate Neural Audio Codec")
|
20 |
st.write("""
|
21 |
-
Upload your audio file, adjust the codec parameters, and compare the original and reconstructed audio.
|
22 |
SemantiCodec achieves high-quality audio reconstruction with ultra-low bitrates!
|
23 |
""")
|
24 |
|
@@ -34,7 +35,7 @@ ddim_steps = st.sidebar.slider("DDIM Sampling Steps", 10, 100, 50, step=5)
|
|
34 |
guidance_scale = st.sidebar.slider("CFG Guidance Scale", 0.5, 5.0, 2.0, step=0.1)
|
35 |
|
36 |
# Upload Audio File
|
37 |
-
uploaded_file = st.file_uploader("Upload an audio file (WAV format)", type=["wav"])
|
38 |
|
39 |
# Helper function: Plot spectrogram
|
40 |
def plot_spectrogram(waveform, sample_rate, title):
|
@@ -57,7 +58,7 @@ if uploaded_file and st.button("Run SemantiCodec"):
|
|
57 |
|
58 |
# Load audio
|
59 |
waveform, sample_rate = torchaudio.load(input_path)
|
60 |
-
|
61 |
# Check if resampling is needed
|
62 |
if sample_rate != DEFAULT_SAMPLE_RATE:
|
63 |
st.write(f"Resampling audio from {sample_rate} Hz to {DEFAULT_SAMPLE_RATE} Hz...")
|
@@ -65,12 +66,23 @@ if uploaded_file and st.button("Run SemantiCodec"):
|
|
65 |
waveform = resampler(waveform)
|
66 |
sample_rate = DEFAULT_SAMPLE_RATE # Update sample rate to 16kHz
|
67 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
68 |
# Convert to numpy for librosa compatibility
|
69 |
-
|
70 |
|
71 |
-
# Plot Original Spectrogram (16kHz resampled)
|
72 |
-
st.write("Original Audio Spectrogram (Resampled to
|
73 |
-
plot_spectrogram(
|
|
|
|
|
|
|
|
|
74 |
|
75 |
# Initialize SemantiCodec
|
76 |
st.write("Initializing SemantiCodec...")
|
@@ -86,7 +98,7 @@ if uploaded_file and st.button("Run SemantiCodec"):
|
|
86 |
|
87 |
# Encode and Decode
|
88 |
st.write("Encoding and Decoding Audio...")
|
89 |
-
tokens = semanticodec.encode(
|
90 |
reconstructed_waveform = semanticodec.decode(tokens)[0, 0]
|
91 |
|
92 |
# Save reconstructed audio
|
@@ -101,8 +113,8 @@ if uploaded_file and st.button("Run SemantiCodec"):
|
|
101 |
st.write(f"Shape of Latent Code: {tokens.shape}")
|
102 |
|
103 |
# Audio Players
|
104 |
-
st.audio(
|
105 |
-
st.write("Original Audio")
|
106 |
st.audio(reconstructed_path, format="audio/wav")
|
107 |
st.write("Reconstructed Audio")
|
108 |
|
@@ -113,6 +125,5 @@ if uploaded_file and st.button("Run SemantiCodec"):
|
|
113 |
file_name="reconstructed_audio.wav",
|
114 |
)
|
115 |
|
116 |
-
|
117 |
# Footer
|
118 |
st.write("Built with [Streamlit](https://streamlit.io) and SemantiCodec")
|
|
|
13 |
DEFAULT_TOKEN_RATE = 100
|
14 |
DEFAULT_SEMANTIC_VOCAB_SIZE = 16384
|
15 |
DEFAULT_SAMPLE_RATE = 16000
|
16 |
+
MAX_DURATION_SECONDS = 30 # Maximum allowed duration
|
17 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
18 |
|
19 |
# Title and Description
|
20 |
st.title("SemantiCodec: Ultra-Low Bitrate Neural Audio Codec")
|
21 |
st.write("""
|
22 |
+
Upload your audio file (up to 30 seconds), adjust the codec parameters, and compare the original and reconstructed audio.
|
23 |
SemantiCodec achieves high-quality audio reconstruction with ultra-low bitrates!
|
24 |
""")
|
25 |
|
|
|
35 |
guidance_scale = st.sidebar.slider("CFG Guidance Scale", 0.5, 5.0, 2.0, step=0.1)
|
36 |
|
37 |
# Upload Audio File
|
38 |
+
uploaded_file = st.file_uploader("Upload an audio file (WAV format, up to 30 seconds)", type=["wav"])
|
39 |
|
40 |
# Helper function: Plot spectrogram
|
41 |
def plot_spectrogram(waveform, sample_rate, title):
|
|
|
58 |
|
59 |
# Load audio
|
60 |
waveform, sample_rate = torchaudio.load(input_path)
|
61 |
+
|
62 |
# Check if resampling is needed
|
63 |
if sample_rate != DEFAULT_SAMPLE_RATE:
|
64 |
st.write(f"Resampling audio from {sample_rate} Hz to {DEFAULT_SAMPLE_RATE} Hz...")
|
|
|
66 |
waveform = resampler(waveform)
|
67 |
sample_rate = DEFAULT_SAMPLE_RATE # Update sample rate to 16kHz
|
68 |
|
69 |
+
# Check and limit duration
|
70 |
+
num_samples = waveform.size(1)
|
71 |
+
max_samples = MAX_DURATION_SECONDS * sample_rate # 30 seconds limit
|
72 |
+
if num_samples > max_samples:
|
73 |
+
st.write(f"Truncating audio to the first {MAX_DURATION_SECONDS} seconds...")
|
74 |
+
waveform = waveform[:, :max_samples]
|
75 |
+
|
76 |
# Convert to numpy for librosa compatibility
|
77 |
+
waveform_np = waveform[0].numpy()
|
78 |
|
79 |
+
# Plot Original Spectrogram (16kHz resampled and truncated)
|
80 |
+
st.write(f"Original Audio Spectrogram (Resampled and limited to {MAX_DURATION_SECONDS} seconds):")
|
81 |
+
plot_spectrogram(waveform_np, sample_rate, f"Original Audio Spectrogram (Resampled to {DEFAULT_SAMPLE_RATE} Hz)")
|
82 |
+
|
83 |
+
# Save truncated audio for processing
|
84 |
+
truncated_path = os.path.join(temp_dir, "truncated_input.wav")
|
85 |
+
torchaudio.save(truncated_path, waveform, sample_rate)
|
86 |
|
87 |
# Initialize SemantiCodec
|
88 |
st.write("Initializing SemantiCodec...")
|
|
|
98 |
|
99 |
# Encode and Decode
|
100 |
st.write("Encoding and Decoding Audio...")
|
101 |
+
tokens = semanticodec.encode(truncated_path)
|
102 |
reconstructed_waveform = semanticodec.decode(tokens)[0, 0]
|
103 |
|
104 |
# Save reconstructed audio
|
|
|
113 |
st.write(f"Shape of Latent Code: {tokens.shape}")
|
114 |
|
115 |
# Audio Players
|
116 |
+
st.audio(truncated_path, format="audio/wav")
|
117 |
+
st.write("Original Audio (Truncated)")
|
118 |
st.audio(reconstructed_path, format="audio/wav")
|
119 |
st.write("Reconstructed Audio")
|
120 |
|
|
|
125 |
file_name="reconstructed_audio.wav",
|
126 |
)
|
127 |
|
|
|
128 |
# Footer
|
129 |
st.write("Built with [Streamlit](https://streamlit.io) and SemantiCodec")
|