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
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
|
// Audio processing code, adapted from whisper.cpp
// https://github.com/ggerganov/whisper.cpp
use super::worker;
pub trait Float: num_traits::Float + num_traits::FloatConst + num_traits::NumAssign {}
impl Float for f32 {}
impl Float for f64 {}
// https://github.com/ggerganov/whisper.cpp/blob/4774d2feb01a772a15de81ffc34b34a1f294f020/whisper.cpp#L2357
fn fft<T: Float>(inp: &[T]) -> Vec<T> {
let n = inp.len();
let zero = T::zero();
if n == 1 {
return vec![inp[0], zero];
}
if n % 2 == 1 {
return dft(inp);
}
let mut out = vec![zero; n * 2];
let mut even = vec![];
even.reserve(n / 2);
let mut odd = vec![];
odd.reserve(n / 2);
for (i, &inp) in inp.iter().enumerate() {
if i % 2 == 0 {
even.push(inp)
} else {
odd.push(inp);
}
}
let even_fft = fft(&even);
let odd_fft = fft(&odd);
let two_pi = T::PI() + T::PI();
let n_t = T::from(n).unwrap();
for k in 0..n / 2 {
let k_t = T::from(k).unwrap();
let theta = two_pi * k_t / n_t;
let re = theta.cos();
let im = -theta.sin();
let re_odd = odd_fft[2 * k];
let im_odd = odd_fft[2 * k + 1];
out[2 * k] = even_fft[2 * k] + re * re_odd - im * im_odd;
out[2 * k + 1] = even_fft[2 * k + 1] + re * im_odd + im * re_odd;
out[2 * (k + n / 2)] = even_fft[2 * k] - re * re_odd + im * im_odd;
out[2 * (k + n / 2) + 1] = even_fft[2 * k + 1] - re * im_odd - im * re_odd;
}
out
}
// https://github.com/ggerganov/whisper.cpp/blob/4774d2feb01a772a15de81ffc34b34a1f294f020/whisper.cpp#L2337
fn dft<T: Float>(inp: &[T]) -> Vec<T> {
let zero = T::zero();
let n = inp.len();
let two_pi = T::PI() + T::PI();
let mut out = Vec::new();
out.reserve(2 * n);
let n_t = T::from(n).unwrap();
for k in 0..n {
let k_t = T::from(k).unwrap();
let mut re = zero;
let mut im = zero;
for (j, &inp) in inp.iter().enumerate() {
let j_t = T::from(j).unwrap();
let angle = two_pi * k_t * j_t / n_t;
re += inp * angle.cos();
im -= inp * angle.sin();
}
out.push(re);
out.push(im);
}
out
}
#[allow(clippy::too_many_arguments)]
// https://github.com/ggerganov/whisper.cpp/blob/4774d2feb01a772a15de81ffc34b34a1f294f020/whisper.cpp#L2414
fn log_mel_spectrogram_w<T: Float>(
ith: usize,
hann: &[T],
samples: &[T],
filters: &[T],
fft_size: usize,
fft_step: usize,
speed_up: bool,
n_len: usize,
n_mel: usize,
n_threads: usize,
) -> Vec<T> {
let n_fft = if speed_up {
1 + fft_size / 4
} else {
1 + fft_size / 2
};
let zero = T::zero();
let half = T::from(0.5).unwrap();
let mut fft_in = vec![zero; fft_size];
let mut mel = vec![zero; n_len * n_mel];
for i in (ith..n_len).step_by(n_threads) {
let offset = i * fft_step;
// apply Hanning window
for j in 0..fft_size {
fft_in[j] = if offset + j < samples.len() {
hann[j] * samples[offset + j]
} else {
zero
}
}
// FFT -> mag^2
let mut fft_out: Vec<T> = fft(&fft_in);
for j in 0..fft_size {
fft_out[j] = fft_out[2 * j] * fft_out[2 * j] + fft_out[2 * j + 1] * fft_out[2 * j + 1];
}
for j in 1..fft_size / 2 {
let v = fft_out[fft_size - j];
fft_out[j] += v;
}
if speed_up {
// scale down in the frequency domain results in a speed up in the time domain
for j in 0..n_fft {
fft_out[j] = half * (fft_out[2 * j] + fft_out[2 * j + 1]);
}
}
// mel spectrogram
for j in 0..n_mel {
let mut sum = zero;
for k in 0..n_fft {
sum += fft_out[k] * filters[j * n_fft + k];
}
mel[j * n_len + i] = T::max(sum, T::from(1e-10).unwrap()).log10();
}
}
mel
}
fn log_mel_spectrogram_<T: Float + std::fmt::Display>(
samples: &[T],
filters: &[T],
fft_size: usize,
fft_step: usize,
n_mel: usize,
speed_up: bool,
) -> Vec<T> {
let zero = T::zero();
let two_pi = T::PI() + T::PI();
let half = T::from(0.5).unwrap();
let one = T::from(1.0).unwrap();
let four = T::from(4.0).unwrap();
let fft_size_t = T::from(fft_size).unwrap();
let hann: Vec<T> = (0..fft_size)
.map(|i| half * (one - ((two_pi * T::from(i).unwrap()) / fft_size_t).cos()))
.collect();
let n_len = samples.len() / fft_step;
// pad audio with at least one extra chunk of zeros
let pad = 100 * worker::CHUNK_LENGTH / 2;
let n_len = if n_len % pad != 0 {
(n_len / pad + 1) * pad
} else {
n_len
};
let n_len = n_len + pad;
let samples = {
let mut samples_padded = samples.to_vec();
let to_add = n_len * fft_step - samples.len();
samples_padded.extend(std::iter::repeat(zero).take(to_add));
samples_padded
};
// Use a single thread for now.
let mut mel = log_mel_spectrogram_w(
0, &hann, &samples, filters, fft_size, fft_step, speed_up, n_len, n_mel, 1,
);
let mmax = mel
.iter()
.max_by(|&u, &v| u.partial_cmp(v).unwrap_or(std::cmp::Ordering::Greater))
.copied()
.unwrap_or(zero)
- T::from(8).unwrap();
for m in mel.iter_mut() {
let v = T::max(*m, mmax);
*m = v / four + one
}
mel
}
pub fn pcm_to_mel<T: Float + std::fmt::Display>(
samples: &[T],
filters: &[T],
) -> anyhow::Result<Vec<T>> {
let mel = log_mel_spectrogram_(
samples,
filters,
worker::N_FFT,
worker::HOP_LENGTH,
worker::N_MELS,
false,
);
Ok(mel)
}
|
