File size: 6,346 Bytes
bc3e180
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
import os
import glob
import yaml
import torch
import __main__
import numpy as np
import soundfile as sf
import librosa
import librosa.display
import matplotlib.pyplot as plt

import gradio as gr

from src.model.nn.synthesizer import Synthesizer
from src.utils.misc import triangular, downsample
from src.utils.plot import state_video as plot_state_video
from src.utils.audio import mel_basis, state_to_wav
from src.utils.control import vibrato as control_vibrato

class ConfigArgument:
    def __getitem__(self,key):
        return getattr(self, key)
    def __setitem__(self,key,value):
        return setattr(self, key, value)
setattr(__main__, "ConfigArgument", ConfigArgument)

def filter_state_dict(ckpt):
    out_dict = {}
    for key in ckpt.keys():
        new_key = key[6:] if str(key)[:6] == 'model.' else key
        out_dict[new_key] = ckpt[key]
    return out_dict

def flush(directory):
    os.makedirs(directory, exist_ok=True)
    files = glob.glob(f'{directory}/*')
    for f in files:
        os.remove(f)

def add_glissando(f_0, Nt, sr, glissando, max_t):
    front = int(0.2 * np.random.rand() * sr * max_t)
    rear =  int((0.2 * np.random.rand() + 0.3) * sr * max_t)
    middle = max(0, len(f_0) - front - rear)
    ramp = glissando * torch.cat((torch.zeros(front), torch.linspace(0,1,middle), torch.ones(rear)), dim=-1)
    return f_0 * (1 + ramp)

def plot_spectrogram(path, x, n_fft=2048, hop_length=512, n_mel=256, samplerate=48000, max_duration=1):
    x_wave = np.zeros(int(max_duration * samplerate))
    x_wave[:len(x)] += x
    x_spec = librosa.stft(
        x_wave, n_fft=n_fft, hop_length=hop_length, win_length=n_fft, pad_mode='reflect')
    mag = np.abs(x_spec) # (n_frames, n_freq)
    mel_fbank = mel_basis(samplerate, n_fft, n_mel) # (n_mel, n_freq)
    mel = np.einsum('ij,jk->ik', mel_fbank, mag) # (n_frames, n_mel)
    
    plt.figure(figsize=(7,7))
    librosa.display.specshow(mel)
    plt.xticks([])
    plt.yticks([])
    plt.clim([0, 30])
    plt.tight_layout()
    plt.savefig(path, transparent=True)
    plt.close('all')
    plt.clf()

with open("ckpt/config.yaml") as stream:
    configs = yaml.safe_load(stream)

with open("ckpt/pitch.yaml") as stream:
    pitch_dict = yaml.safe_load(stream)

def get_data(duration, resolution, note, glissando, vibrato, stiffness, tension, pluck, amplitude):
    sr = configs['sr']
    Nt = int(duration * sr)
    Nx = int(resolution)

    xgrid = torch.linspace(0,1,Nx)
    tgrid = torch.arange(Nt) / sr
    pitch = pitch_dict[note]
  
    t60_min_1=20.; t60_max_1=30.; t60_min_2=30.; t60_max_2=30.
    t60_diff_max=5.
    T60 = torch.Tensor([[[1000., 25.],[100., 30.]]])

    Nw = int(Nt / configs['block_size']) + 1
    
    xg, tg = torch.meshgrid(xgrid, tgrid, indexing='ij')
    ka = torch.Tensor([stiffness]).view(-1,1) # (1,1)
    al = torch.Tensor([tension]).view(-1,1) # (1,1)
    f_0 = torch.ones(Nt) * pitch # (Nt)
    nx  = torch.Tensor([[[Nx]]]).float()
    p_x = torch.ones_like(nx) * pluck
    p_a = torch.ones_like(nx) * amplitude
    u_0 = triangular(Nx, nx, p_x, p_a) # (1, 1, Nx)
   
    f_0 = add_glissando(f_0, Nt, sr, glissando, Nt / sr)
    f_0 = f_0 + control_vibrato(f_0.view(1,-1), 1/sr, mf=[3.,5.], ma=vibrato)
    f_0 = downsample(f_0, factor=configs['block_size'])

    xg  = xg[:,0].view(-1,1) # (Nx, 1)
    tg  = tg                 # (Nx, Nt)
    ka  = ka.repeat(Nx,1)    # (Nx, 1)
    al  = al.repeat(Nx,1)    # (Nx, 1)
    T60 = T60                # (Nx, 1, 1)
    f_0 = f_0.repeat(Nx,1)   # (Nx, Nw)
    u_0 = u_0.repeat(Nx,1,1) # (Nx, 1, Nx)
    
    params = [xg, tg, ka, al, T60, None, None]
    return params, f_0, u_0
 
def run(duration, resolution, pitch, glissando, vibrato, stiffness, tension, pluck, amplitude):
    checkpoint = torch.load('ckpt/dmsp.ckpt', map_location='cpu')
    checkpoint = filter_state_dict(checkpoint['state_dict'])
    model = Synthesizer(**configs)
    model.load_state_dict(checkpoint)
    
    params, f_0, u_0 = get_data( \
        duration, resolution, pitch, glissando, vibrato, stiffness, tension, pluck, amplitude)

    with torch.no_grad():
        ut, mode_input, mode_output = model(params, f_0, u_0)
    ut = ut.detach() # (Nx, Nt)
    ut_wave = configs['gain'] * ut.mean(0)
    
    save_dir = 'results'
    prefix = 'dmsp'
    fname = 'output'
    flush(save_dir)
    audio_name = f'{save_dir}/{fname}.wav'
    video_name = f'{save_dir}/{prefix}-{fname}.mp4'
    spec_name  = f'{save_dir}/spec.png'

    ut = ut.numpy().T
    ut_wave = ut_wave.numpy()
    maxy = 0.022
    sf.write(audio_name, ut_wave, samplerate=configs['sr'])
    plot_spectrogram(spec_name, ut_wave, samplerate=configs['sr'])
    plot_state_video(save_dir, ut, configs['sr'], prefix=prefix, fname=fname, maxy=maxy)
    return spec_name, video_name

pitch_list = ["G2", "Ab2", "A2", "Bb2", "B2", "C3", "Db3", "D3", "Eb3", "E3", "F3", "Gb3", "G3", "Ab3", "A3", "Bb3", "B3", "C4", "Db4", "D4", "Eb4", "E4", "F4", "Gb4", "G4",]

duration   = gr.Slider(0.1, 1.0, value=1.0, label="Time Duration")
resolution = gr.Slider(128, 256, value=256, label="Space Resolution", info='Reduce to simulate faster. Recommended to leave it as 256.')
pitch      = gr.Dropdown(pitch_list, value="C3", label="Pitch", info="Specify the fundamental frequency as a musical note.")
glissando  = gr.Slider(-0.4, 0.4, value=0, label="Glissando", info='Set +/- to ascend (+) or descend (-) the pitch')
vibrato    = gr.Slider(0, 0.25, value=0, label="Vibrato", info='Set larger value to add more vibrato')
stiffness  = gr.Slider(0.011, 0.029, value=0.02, label="Stiffness", info='Stiffness can change the resulting pitch. Specify low values when tension is high')
tension    = gr.Slider(1.0,  25, value=4, label="Tension", info='Tension can introduce non-linear effects such as pitch glide. Specify low values when stiffness is high')
pluck      = gr.Slider(0.12, 0.5, value=0.2, label="Pluck Position", info='Peak position of an initial condition')
amplitude  = gr.Slider(0.001, 0.02, value=0.015, label="Pluck Amplitude", info='Peak amplitude of an initial condition')

demo = gr.Interface(
    fn=run,
    inputs=[
        duration, resolution, pitch, glissando, vibrato,
        stiffness, tension, pluck, amplitude,
    ],
    outputs=[
        gr.Image(),
        gr.Video(format='mp4', include_audio=True),
    ],
)
demo.launch()