Spaces:
Runtime error
Runtime error
Commit
•
02063a3
1
Parent(s):
8ff0f67
Create spectro.py
Browse files- spectro.py +207 -0
spectro.py
ADDED
@@ -0,0 +1,207 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Audio processing tools to convert between spectrogram images and waveforms.
|
3 |
+
"""
|
4 |
+
import io
|
5 |
+
import typing as T
|
6 |
+
|
7 |
+
import numpy as np
|
8 |
+
from PIL import Image
|
9 |
+
import pydub
|
10 |
+
from scipy.io import wavfile
|
11 |
+
import torch
|
12 |
+
import torchaudio
|
13 |
+
|
14 |
+
|
15 |
+
def wav_bytes_from_spectrogram_image(image: Image.Image) -> T.Tuple[io.BytesIO, float]:
|
16 |
+
"""
|
17 |
+
Reconstruct a WAV audio clip from a spectrogram image. Also returns the duration in seconds.
|
18 |
+
"""
|
19 |
+
|
20 |
+
max_volume = 50
|
21 |
+
power_for_image = 0.25
|
22 |
+
Sxx = spectrogram_from_image(image, max_volume=max_volume, power_for_image=power_for_image)
|
23 |
+
|
24 |
+
sample_rate = 44100 # [Hz]
|
25 |
+
clip_duration_ms = 5000 # [ms]
|
26 |
+
|
27 |
+
bins_per_image = 512
|
28 |
+
n_mels = 512
|
29 |
+
|
30 |
+
# FFT parameters
|
31 |
+
window_duration_ms = 100 # [ms]
|
32 |
+
padded_duration_ms = 400 # [ms]
|
33 |
+
step_size_ms = 10 # [ms]
|
34 |
+
|
35 |
+
# Derived parameters
|
36 |
+
num_samples = int(image.width / float(bins_per_image) * clip_duration_ms) * sample_rate
|
37 |
+
n_fft = int(padded_duration_ms / 1000.0 * sample_rate)
|
38 |
+
hop_length = int(step_size_ms / 1000.0 * sample_rate)
|
39 |
+
win_length = int(window_duration_ms / 1000.0 * sample_rate)
|
40 |
+
|
41 |
+
samples = waveform_from_spectrogram(
|
42 |
+
Sxx=Sxx,
|
43 |
+
n_fft=n_fft,
|
44 |
+
hop_length=hop_length,
|
45 |
+
win_length=win_length,
|
46 |
+
num_samples=num_samples,
|
47 |
+
sample_rate=sample_rate,
|
48 |
+
mel_scale=True,
|
49 |
+
n_mels=n_mels,
|
50 |
+
max_mel_iters=200,
|
51 |
+
num_griffin_lim_iters=32,
|
52 |
+
)
|
53 |
+
|
54 |
+
wav_bytes = io.BytesIO()
|
55 |
+
wavfile.write(wav_bytes, sample_rate, samples.astype(np.int16))
|
56 |
+
wav_bytes.seek(0)
|
57 |
+
|
58 |
+
duration_s = float(len(samples)) / sample_rate
|
59 |
+
|
60 |
+
return wav_bytes, duration_s
|
61 |
+
|
62 |
+
|
63 |
+
def spectrogram_from_image(
|
64 |
+
image: Image.Image, max_volume: float = 50, power_for_image: float = 0.25
|
65 |
+
) -> np.ndarray:
|
66 |
+
"""
|
67 |
+
Compute a spectrogram magnitude array from a spectrogram image.
|
68 |
+
TODO(hayk): Add image_from_spectrogram and call this out as the reverse.
|
69 |
+
"""
|
70 |
+
# Convert to a numpy array of floats
|
71 |
+
data = np.array(image).astype(np.float32)
|
72 |
+
|
73 |
+
# Flip Y take a single channel
|
74 |
+
data = data[::-1, :, 0]
|
75 |
+
|
76 |
+
# Invert
|
77 |
+
data = 255 - data
|
78 |
+
|
79 |
+
# Rescale to max volume
|
80 |
+
data = data * max_volume / 255
|
81 |
+
|
82 |
+
# Reverse the power curve
|
83 |
+
data = np.power(data, 1 / power_for_image)
|
84 |
+
|
85 |
+
return data
|
86 |
+
|
87 |
+
|
88 |
+
def spectrogram_from_waveform(
|
89 |
+
waveform: np.ndarray,
|
90 |
+
sample_rate: int,
|
91 |
+
n_fft: int,
|
92 |
+
hop_length: int,
|
93 |
+
win_length: int,
|
94 |
+
mel_scale: bool = True,
|
95 |
+
n_mels: int = 512,
|
96 |
+
) -> np.ndarray:
|
97 |
+
"""
|
98 |
+
Compute a spectrogram from a waveform.
|
99 |
+
"""
|
100 |
+
|
101 |
+
spectrogram_func = torchaudio.transforms.Spectrogram(
|
102 |
+
n_fft=n_fft,
|
103 |
+
power=None,
|
104 |
+
hop_length=hop_length,
|
105 |
+
win_length=win_length,
|
106 |
+
)
|
107 |
+
|
108 |
+
waveform_tensor = torch.from_numpy(waveform.astype(np.float32)).reshape(1, -1)
|
109 |
+
Sxx_complex = spectrogram_func(waveform_tensor).numpy()[0]
|
110 |
+
|
111 |
+
Sxx_mag = np.abs(Sxx_complex)
|
112 |
+
|
113 |
+
if mel_scale:
|
114 |
+
mel_scaler = torchaudio.transforms.MelScale(
|
115 |
+
n_mels=n_mels,
|
116 |
+
sample_rate=sample_rate,
|
117 |
+
f_min=0,
|
118 |
+
f_max=10000,
|
119 |
+
n_stft=n_fft // 2 + 1,
|
120 |
+
norm=None,
|
121 |
+
mel_scale="htk",
|
122 |
+
)
|
123 |
+
|
124 |
+
Sxx_mag = mel_scaler(torch.from_numpy(Sxx_mag)).numpy()
|
125 |
+
|
126 |
+
return Sxx_mag
|
127 |
+
|
128 |
+
|
129 |
+
def waveform_from_spectrogram(
|
130 |
+
Sxx: np.ndarray,
|
131 |
+
n_fft: int,
|
132 |
+
hop_length: int,
|
133 |
+
win_length: int,
|
134 |
+
num_samples: int,
|
135 |
+
sample_rate: int,
|
136 |
+
mel_scale: bool = True,
|
137 |
+
n_mels: int = 512,
|
138 |
+
max_mel_iters: int = 200,
|
139 |
+
num_griffin_lim_iters: int = 32,
|
140 |
+
device: str = "cuda:0",
|
141 |
+
) -> np.ndarray:
|
142 |
+
"""
|
143 |
+
Reconstruct a waveform from a spectrogram.
|
144 |
+
This is an approximate inverse of spectrogram_from_waveform, using the Griffin-Lim algorithm
|
145 |
+
to approximate the phase.
|
146 |
+
"""
|
147 |
+
Sxx_torch = torch.from_numpy(Sxx).to(device)
|
148 |
+
|
149 |
+
# TODO(hayk): Make this a class that caches the two things
|
150 |
+
|
151 |
+
if mel_scale:
|
152 |
+
mel_inv_scaler = torchaudio.transforms.InverseMelScale(
|
153 |
+
n_mels=n_mels,
|
154 |
+
sample_rate=sample_rate,
|
155 |
+
f_min=0,
|
156 |
+
f_max=10000,
|
157 |
+
n_stft=n_fft // 2 + 1,
|
158 |
+
norm=None,
|
159 |
+
mel_scale="htk",
|
160 |
+
max_iter=max_mel_iters,
|
161 |
+
).to(device)
|
162 |
+
|
163 |
+
Sxx_torch = mel_inv_scaler(Sxx_torch)
|
164 |
+
|
165 |
+
griffin_lim = torchaudio.transforms.GriffinLim(
|
166 |
+
n_fft=n_fft,
|
167 |
+
win_length=win_length,
|
168 |
+
hop_length=hop_length,
|
169 |
+
power=1.0,
|
170 |
+
n_iter=num_griffin_lim_iters,
|
171 |
+
).to(device)
|
172 |
+
|
173 |
+
waveform = griffin_lim(Sxx_torch).cpu().numpy()
|
174 |
+
|
175 |
+
return waveform
|
176 |
+
|
177 |
+
|
178 |
+
def mp3_bytes_from_wav_bytes(wav_bytes: io.BytesIO) -> io.BytesIO:
|
179 |
+
mp3_bytes = io.BytesIO()
|
180 |
+
sound = pydub.AudioSegment.from_wav(wav_bytes)
|
181 |
+
sound.export(mp3_bytes, format="mp3")
|
182 |
+
mp3_bytes.seek(0)
|
183 |
+
return mp3_bytes
|
184 |
+
|
185 |
+
def image_from_spectrogram(spectrogram: np.ndarray, max_volume: float = 50, power_for_image: float = 0.25) -> Image.Image:
|
186 |
+
"""
|
187 |
+
Compute a spectrogram image from a spectrogram magnitude array.
|
188 |
+
"""
|
189 |
+
# Apply the power curve
|
190 |
+
data = np.power(spectrogram, power_for_image)
|
191 |
+
|
192 |
+
# Rescale to 0-255
|
193 |
+
data = data * 255 / max_volume
|
194 |
+
|
195 |
+
# Invert
|
196 |
+
data = 255 - data
|
197 |
+
|
198 |
+
# Convert to a PIL image
|
199 |
+
image = Image.fromarray(data.astype(np.uint8))
|
200 |
+
|
201 |
+
# Flip Y
|
202 |
+
image = image.transpose(Image.FLIP_TOP_BOTTOM)
|
203 |
+
|
204 |
+
# Convert to RGB
|
205 |
+
image = image.convert("RGB")
|
206 |
+
|
207 |
+
return image
|