File size: 8,614 Bytes
2bbfbb7 0393dfa 2bbfbb7 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 |
//! BigVGAN vocoder implementation
//!
//! High-quality neural vocoder for mel-spectrogram to waveform conversion
use crate::{Error, Result};
use ndarray::{Array2, IxDyn};
use std::collections::HashMap;
use std::path::Path;
use crate::model::OnnxSession;
use super::{Vocoder, snake_activation_vec};
/// BigVGAN configuration
#[derive(Debug, Clone)]
pub struct BigVGANConfig {
/// Sample rate
pub sample_rate: u32,
/// Number of mel channels
pub num_mels: usize,
/// Upsampling rates
pub upsample_rates: Vec<usize>,
/// Upsampling kernel sizes
pub upsample_kernel_sizes: Vec<usize>,
/// ResBlock kernel sizes
pub resblock_kernel_sizes: Vec<usize>,
/// ResBlock dilation sizes
pub resblock_dilation_sizes: Vec<Vec<usize>>,
/// Initial channel size
pub upsample_initial_channel: usize,
/// Use anti-aliasing
pub use_anti_alias: bool,
}
impl Default for BigVGANConfig {
fn default() -> Self {
Self {
sample_rate: 22050,
num_mels: 80,
upsample_rates: vec![8, 8, 2, 2],
upsample_kernel_sizes: vec![16, 16, 4, 4],
resblock_kernel_sizes: vec![3, 7, 11],
resblock_dilation_sizes: vec![vec![1, 3, 5], vec![1, 3, 5], vec![1, 3, 5]],
upsample_initial_channel: 512,
use_anti_alias: true,
}
}
}
impl BigVGANConfig {
/// Calculate total upsampling factor
pub fn total_upsample_factor(&self) -> usize {
self.upsample_rates.iter().product()
}
/// Get hop length (same as upsample factor)
pub fn hop_length(&self) -> usize {
self.total_upsample_factor()
}
}
/// BigVGAN vocoder
pub struct BigVGAN {
session: Option<OnnxSession>,
config: BigVGANConfig,
}
impl BigVGAN {
/// Load BigVGAN from ONNX model
pub fn load<P: AsRef<Path>>(path: P, config: BigVGANConfig) -> Result<Self> {
let session = OnnxSession::load(path)?;
Ok(Self {
session: Some(session),
config,
})
}
/// Create BigVGAN with fallback synthesizer
pub fn new_fallback(config: BigVGANConfig) -> Self {
Self {
session: None,
config,
}
}
/// Get configuration
pub fn config(&self) -> &BigVGANConfig {
&self.config
}
/// Synthesize audio using fallback algorithm
fn synthesize_fallback(&self, mel: &Array2<f32>) -> Result<Vec<f32>> {
// Simple overlap-add synthesis as fallback
let num_frames = mel.ncols();
let hop_length = self.config.hop_length();
let frame_size = hop_length * 4; // Use 4x overlap
let output_length = (num_frames - 1) * hop_length + frame_size;
let mut output = vec![0.0f32; output_length];
let mut window_sum = vec![0.0f32; output_length];
// Hann window
let window: Vec<f32> = (0..frame_size)
.map(|n| {
0.5 * (1.0 - (2.0 * std::f32::consts::PI * n as f32 / frame_size as f32).cos())
})
.collect();
// Generate frames from mel
for frame_idx in 0..num_frames {
let start = frame_idx * hop_length;
// Generate frame from mel (simplified: use mel features to modulate noise)
let mel_frame: Vec<f32> = (0..self.config.num_mels)
.map(|i| mel[[i, frame_idx]])
.collect();
// Generate frame using mel features
let frame = self.generate_frame(&mel_frame, frame_size);
// Overlap-add
for i in 0..frame_size {
if start + i < output_length {
output[start + i] += frame[i] * window[i];
window_sum[start + i] += window[i] * window[i];
}
}
}
// Normalize by window sum
for i in 0..output_length {
if window_sum[i] > 1e-8 {
output[i] /= window_sum[i];
}
}
// Apply post-processing
let output = snake_activation_vec(&output, 0.3);
Ok(output)
}
/// Generate a single frame from mel features
fn generate_frame(&self, mel: &[f32], frame_size: usize) -> Vec<f32> {
use rand::Rng;
let mut rng = rand::thread_rng();
// Compute overall energy from mel
let energy: f32 = mel.iter().map(|x| x.exp()).sum::<f32>() / mel.len() as f32;
let energy = energy.sqrt().min(2.0);
// Generate frame with harmonic content
let mut frame = vec![0.0f32; frame_size];
// Use mel bands to create frequency content
for (freq_idx, &mel_val) in mel.iter().enumerate() {
let freq = (freq_idx as f32 / mel.len() as f32) * (self.config.sample_rate as f32 / 2.0);
let amplitude = mel_val.exp().min(1.0) * 0.1;
// Add harmonic
for i in 0..frame_size {
let t = i as f32 / self.config.sample_rate as f32;
frame[i] += amplitude * (2.0 * std::f32::consts::PI * freq * t).sin();
}
}
// Add filtered noise
for i in 0..frame_size {
frame[i] += rng.gen_range(-0.1..0.1) * energy * 0.1;
}
// Normalize
let max_abs = frame.iter().map(|x| x.abs()).fold(0.0f32, f32::max);
if max_abs > 1.0 {
for v in frame.iter_mut() {
*v /= max_abs;
}
}
frame
}
/// Apply post-processing to output
pub fn post_process(&self, audio: &[f32]) -> Vec<f32> {
use crate::audio::{normalize_audio, apply_fade};
let normalized = normalize_audio(audio);
// Apply fade to avoid clicks
let fade_samples = (self.config.sample_rate as f32 * 0.01) as usize; // 10ms fade
apply_fade(&normalized, fade_samples, fade_samples)
}
}
impl Vocoder for BigVGAN {
fn synthesize(&self, mel: &Array2<f32>) -> Result<Vec<f32>> {
if let Some(ref session) = self.session {
// Use ONNX model
let input = mel.clone().into_shape(IxDyn(&[1, mel.nrows(), mel.ncols()]))?;
let mut inputs = HashMap::new();
inputs.insert("mel".to_string(), input);
let outputs = session.run(inputs)?;
let audio = outputs
.get("audio")
.ok_or_else(|| Error::Vocoder("Missing audio output".into()))?;
// Extract audio samples
let samples: Vec<f32> = audio.iter().cloned().collect();
Ok(self.post_process(&samples))
} else {
// Use fallback synthesis
let audio = self.synthesize_fallback(mel)?;
Ok(self.post_process(&audio))
}
}
fn sample_rate(&self) -> u32 {
self.config.sample_rate
}
fn hop_length(&self) -> usize {
self.config.hop_length()
}
}
/// Helper function to create BigVGAN for 22kHz audio
pub fn create_bigvgan_22k() -> BigVGAN {
let config = BigVGANConfig {
sample_rate: 22050,
..Default::default()
};
BigVGAN::new_fallback(config)
}
/// Helper function to create BigVGAN for 24kHz audio
pub fn create_bigvgan_24k() -> BigVGAN {
let config = BigVGANConfig {
sample_rate: 24000,
upsample_rates: vec![12, 10, 2, 2],
..Default::default()
};
BigVGAN::new_fallback(config)
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_bigvgan_config() {
let config = BigVGANConfig::default();
assert_eq!(config.total_upsample_factor(), 256);
assert_eq!(config.hop_length(), 256);
}
#[test]
fn test_bigvgan_fallback() {
let vocoder = create_bigvgan_22k();
assert_eq!(vocoder.sample_rate(), 22050);
// Create small test mel
let mel = Array2::zeros((80, 10));
let result = vocoder.synthesize(&mel);
assert!(result.is_ok());
let audio = result.unwrap();
assert!(audio.len() > 0);
}
#[test]
fn test_generate_frame() {
let vocoder = create_bigvgan_22k();
let mel = vec![0.0f32; 80];
let frame = vocoder.generate_frame(&mel, 256);
assert_eq!(frame.len(), 256);
}
#[test]
fn test_post_process() {
let vocoder = create_bigvgan_22k();
let audio = vec![0.5f32; 1000];
let processed = vocoder.post_process(&audio);
assert_eq!(processed.len(), audio.len());
// Check fade was applied (first samples should be smaller)
assert!(processed[0].abs() < 0.1);
}
}
|