Spaces:
Running
on
T4
Running
on
T4
Upload 2 files
Browse files- app.py +118 -0
- requirements.txt +6 -0
app.py
ADDED
@@ -0,0 +1,118 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
import torchaudio
|
3 |
+
import torch
|
4 |
+
import librosa
|
5 |
+
import librosa.display
|
6 |
+
import matplotlib.pyplot as plt
|
7 |
+
from semanticodec import SemantiCodec
|
8 |
+
import numpy as np
|
9 |
+
import tempfile
|
10 |
+
import os
|
11 |
+
|
12 |
+
# Set default parameters
|
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 |
+
|
25 |
+
# Sidebar: Parameters
|
26 |
+
st.sidebar.title("Codec Parameters")
|
27 |
+
token_rate = st.sidebar.selectbox("Token Rate (tokens/sec)", [25, 50, 100], index=2)
|
28 |
+
semantic_vocab_size = st.sidebar.selectbox(
|
29 |
+
"Semantic Vocabulary Size",
|
30 |
+
[4096, 8192, 16384, 32768],
|
31 |
+
index=2,
|
32 |
+
)
|
33 |
+
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):
|
41 |
+
plt.figure(figsize=(10, 4))
|
42 |
+
S = librosa.feature.melspectrogram(y=waveform, sr=sample_rate, n_mels=128, fmax=sample_rate // 2)
|
43 |
+
S_dB = librosa.power_to_db(S, ref=np.max)
|
44 |
+
librosa.display.specshow(S_dB, sr=sample_rate, x_axis='time', y_axis='mel', cmap='viridis')
|
45 |
+
plt.colorbar(format='%+2.0f dB')
|
46 |
+
plt.title(title)
|
47 |
+
plt.tight_layout()
|
48 |
+
st.pyplot(plt)
|
49 |
+
|
50 |
+
# Process Audio
|
51 |
+
if uploaded_file and st.button("Run SemantiCodec"):
|
52 |
+
with tempfile.TemporaryDirectory() as temp_dir:
|
53 |
+
# Save uploaded file
|
54 |
+
input_path = os.path.join(temp_dir, "input.wav")
|
55 |
+
with open(input_path, "wb") as f:
|
56 |
+
f.write(uploaded_file.read())
|
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...")
|
64 |
+
resampler = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=DEFAULT_SAMPLE_RATE)
|
65 |
+
waveform = resampler(waveform)
|
66 |
+
sample_rate = DEFAULT_SAMPLE_RATE # Update sample rate to 16kHz
|
67 |
+
|
68 |
+
# Convert to numpy for librosa compatibility
|
69 |
+
waveform = waveform[0].numpy()
|
70 |
+
|
71 |
+
# Plot Original Spectrogram (16kHz resampled)
|
72 |
+
st.write("Original Audio Spectrogram (Resampled to 16kHz):")
|
73 |
+
plot_spectrogram(waveform, sample_rate, "Original Audio Spectrogram (Resampled to 16kHz)")
|
74 |
+
|
75 |
+
# Initialize SemantiCodec
|
76 |
+
st.write("Initializing SemantiCodec...")
|
77 |
+
semanticodec = SemantiCodec(
|
78 |
+
token_rate=token_rate,
|
79 |
+
semantic_vocab_size=semantic_vocab_size,
|
80 |
+
ddim_sample_step=ddim_steps,
|
81 |
+
cfg_scale=guidance_scale,
|
82 |
+
)
|
83 |
+
semanticodec.device = device
|
84 |
+
semanticodec.encoder = semanticodec.encoder.to(device)
|
85 |
+
semanticodec.decoder = semanticodec.decoder.to(device)
|
86 |
+
|
87 |
+
# Encode and Decode
|
88 |
+
st.write("Encoding and Decoding Audio...")
|
89 |
+
tokens = semanticodec.encode(input_path)
|
90 |
+
reconstructed_waveform = semanticodec.decode(tokens)[0, 0]
|
91 |
+
|
92 |
+
# Save reconstructed audio
|
93 |
+
reconstructed_path = os.path.join(temp_dir, "reconstructed.wav")
|
94 |
+
torchaudio.save(reconstructed_path, torch.tensor([reconstructed_waveform]), sample_rate)
|
95 |
+
|
96 |
+
# Plot Reconstructed Spectrogram
|
97 |
+
st.write("Reconstructed Audio Spectrogram:")
|
98 |
+
plot_spectrogram(reconstructed_waveform, sample_rate, "Reconstructed Audio Spectrogram")
|
99 |
+
|
100 |
+
# Display latent code shape
|
101 |
+
st.write(f"Shape of Latent Code: {tokens.shape}")
|
102 |
+
|
103 |
+
# Audio Players
|
104 |
+
st.audio(input_path, format="audio/wav")
|
105 |
+
st.write("Original Audio")
|
106 |
+
st.audio(reconstructed_path, format="audio/wav")
|
107 |
+
st.write("Reconstructed Audio")
|
108 |
+
|
109 |
+
# Download Button for Reconstructed Audio
|
110 |
+
st.download_button(
|
111 |
+
"Download Reconstructed Audio",
|
112 |
+
data=open(reconstructed_path, "rb").read(),
|
113 |
+
file_name="reconstructed_audio.wav",
|
114 |
+
)
|
115 |
+
|
116 |
+
|
117 |
+
# Footer
|
118 |
+
st.write("Built with [Streamlit](https://streamlit.io) and SemantiCodec")
|
requirements.txt
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
git+https://github.com/haoheliu/SemantiCodec-inference.git
|
2 |
+
matplotlib
|
3 |
+
librosa
|
4 |
+
torch
|
5 |
+
torchaudio
|
6 |
+
streamlit
|