File size: 10,193 Bytes
97e3499 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 | """
Phase 7: Evaluation & Comparison
Generates quantitative metrics and comparison charts for all three
model stages: baseline, improved pipeline, and fine-tuned.
Metrics:
- Audio duration and pacing analysis
- Signal-to-noise ratio (SNR)
- Real-time factor (RTF)
- Spectral analysis (energy distribution)
- Waveform comparison charts
Output:
docs/images/comparison_waveforms.png
docs/images/comparison_spectrogram.png
docs/images/comparison_duration.png
docs/images/training_loss.png
docs/benchmarks/evaluation.json
Usage:
conda activate new-arabic-tts
python scripts/evaluate.py
"""
import os
import json
import numpy as np
import soundfile as sf
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from pathlib import Path
PROJECT_ROOT = Path(__file__).resolve().parent.parent
IMAGES_DIR = PROJECT_ROOT / "docs" / "images"
BENCHMARKS_DIR = PROJECT_ROOT / "docs" / "benchmarks"
# Output files from each phase
FILES = {
"Phase 1: Baseline": PROJECT_ROOT / "outputs" / "base_model_test.wav",
"Phase 6: Fine-tuned": PROJECT_ROOT / "outputs" / "finetuned_model_test.wav",
}
COLORS = {
"Phase 1: Baseline": "#e74c3c",
"Phase 6: Fine-tuned": "#2ecc71",
}
def measure_audio(wav, sr):
"""Compute audio metrics."""
duration = len(wav) / sr
rms = np.sqrt(np.mean(wav ** 2))
peak = np.max(np.abs(wav))
# SNR
frame_size = int(sr * 0.025)
hop = frame_size // 2
frames = [wav[i:i + frame_size] for i in range(0, len(wav) - frame_size, hop)]
rms_vals = sorted([np.sqrt(np.mean(f ** 2)) for f in frames])
n = len(rms_vals)
noise = np.mean(rms_vals[:max(1, n // 10)])
signal = np.mean(rms_vals[n // 2:])
snr = 20 * np.log10(signal / max(noise, 1e-10))
# Silence ratio
silence = np.mean(np.abs(wav) < 0.01)
# Speaking rate (non-silent frames)
speaking_frames = sum(1 for f in frames if np.sqrt(np.mean(f ** 2)) > 0.01)
speaking_duration = speaking_frames * (hop / sr)
return {
"duration_s": round(duration, 2),
"rms": round(float(rms), 6),
"peak": round(float(peak), 6),
"snr_db": round(float(snr), 1),
"silence_ratio": round(float(silence), 3),
"speaking_duration_s": round(speaking_duration, 2),
}
def plot_waveforms(audio_data, sr_data):
"""Side-by-side waveform comparison."""
n_plots = len(audio_data)
fig, axes = plt.subplots(n_plots, 1, figsize=(14, 4 * n_plots), sharex=False)
if n_plots == 1:
axes = [axes]
for ax, (name, wav) in zip(axes, audio_data.items()):
sr = sr_data[name]
time_axis = np.arange(len(wav)) / sr
color = COLORS.get(name, "#3498db")
ax.plot(time_axis, wav, color=color, linewidth=0.3, alpha=0.8)
ax.fill_between(time_axis, wav, alpha=0.15, color=color)
ax.set_ylabel("Amplitude")
ax.set_title(name, fontsize=11, fontweight="bold", color=color)
ax.set_ylim(-1, 1)
ax.set_xlim(0, max(len(w) / sr_data[n] for n, w in audio_data.items()))
ax.grid(True, alpha=0.2)
ax.axhline(y=0, color="gray", linewidth=0.5)
axes[-1].set_xlabel("Time (seconds)")
fig.suptitle("Waveform Comparison — Base vs Fine-tuned", fontsize=14, fontweight="bold")
fig.tight_layout()
fig.savefig(IMAGES_DIR / "comparison_waveforms.png", dpi=150)
plt.close(fig)
print(" Saved comparison_waveforms.png")
def plot_spectrograms(audio_data, sr_data):
"""Spectrogram comparison."""
n_plots = len(audio_data)
fig, axes = plt.subplots(n_plots, 1, figsize=(14, 4 * n_plots))
if n_plots == 1:
axes = [axes]
for ax, (name, wav) in zip(axes, audio_data.items()):
sr = sr_data[name]
ax.specgram(wav, Fs=sr, NFFT=1024, noverlap=512, cmap="magma")
ax.set_ylabel("Frequency (Hz)")
ax.set_title(name, fontsize=11, fontweight="bold")
ax.set_ylim(0, 8000)
axes[-1].set_xlabel("Time (seconds)")
fig.suptitle("Spectrogram Comparison — Base vs Fine-tuned", fontsize=14, fontweight="bold")
fig.tight_layout()
fig.savefig(IMAGES_DIR / "comparison_spectrogram.png", dpi=150)
plt.close(fig)
print(" Saved comparison_spectrogram.png")
def plot_metrics_comparison(metrics):
"""Bar chart comparing key metrics."""
fig, axes = plt.subplots(1, 4, figsize=(16, 5))
names = list(metrics.keys())
short_names = [n.split(": ")[-1] for n in names]
colors = [COLORS.get(n, "#3498db") for n in names]
# Duration
vals = [metrics[n]["duration_s"] for n in names]
axes[0].bar(short_names, vals, color=colors, edgecolor="white", linewidth=1.5)
axes[0].set_title("Duration (seconds)")
axes[0].set_ylabel("Seconds")
for i, v in enumerate(vals):
axes[0].text(i, v + 0.2, f"{v}s", ha="center", fontsize=10, fontweight="bold")
# SNR
vals = [metrics[n]["snr_db"] for n in names]
axes[1].bar(short_names, vals, color=colors, edgecolor="white", linewidth=1.5)
axes[1].set_title("Signal-to-Noise Ratio")
axes[1].set_ylabel("dB")
for i, v in enumerate(vals):
axes[1].text(i, v + 0.5, f"{v}", ha="center", fontsize=10, fontweight="bold")
# Speaking duration (non-silence)
vals = [metrics[n]["speaking_duration_s"] for n in names]
axes[2].bar(short_names, vals, color=colors, edgecolor="white", linewidth=1.5)
axes[2].set_title("Speaking Duration")
axes[2].set_ylabel("Seconds")
for i, v in enumerate(vals):
axes[2].text(i, v + 0.2, f"{v}s", ha="center", fontsize=10, fontweight="bold")
# Silence ratio
vals = [metrics[n]["silence_ratio"] * 100 for n in names]
axes[3].bar(short_names, vals, color=colors, edgecolor="white", linewidth=1.5)
axes[3].set_title("Silence Ratio")
axes[3].set_ylabel("%")
for i, v in enumerate(vals):
axes[3].text(i, v + 0.5, f"{v:.1f}%", ha="center", fontsize=10, fontweight="bold")
fig.suptitle("Audio Metrics Comparison — Base vs Fine-tuned",
fontsize=14, fontweight="bold")
fig.tight_layout()
fig.savefig(IMAGES_DIR / "comparison_metrics.png", dpi=150)
plt.close(fig)
print(" Saved comparison_metrics.png")
def plot_training_loss():
"""Plot training loss curve from trainer log."""
# Find trainer log
ft_dir = PROJECT_ROOT / "models" / "finetuned" / "run" / "training"
log_files = list(ft_dir.rglob("trainer_0_log.txt"))
if not log_files:
print(" No trainer log found, skipping loss plot")
return
log_path = log_files[0]
steps, losses = [], []
current_step = 0
with open(log_path, "r") as f:
for line in f:
line = line.strip()
# Parse step from: --> TIME: ... -- STEP: 50/1082 -- GLOBAL_STEP: 50
if "GLOBAL_STEP:" in line:
try:
current_step = int(line.split("GLOBAL_STEP:")[1].strip().rstrip("\x1b[0m"))
except (ValueError, IndexError):
pass
# Parse loss from: | > loss: 2.543 (2.543)
if "| > loss:" in line and "loss_text" not in line and "loss_mel" not in line:
try:
val = float(line.split("loss:")[1].strip().split()[0])
if current_step > 0:
steps.append(current_step)
losses.append(val)
except (ValueError, IndexError):
pass
if not steps:
print(" No loss data in log, skipping loss plot")
return
fig, ax = plt.subplots(figsize=(10, 5))
ax.plot(steps, losses, color="#2ecc71", linewidth=1.5, alpha=0.8)
# Smoothed line
if len(losses) > 20:
window = min(50, len(losses) // 5)
smoothed = np.convolve(losses, np.ones(window) / window, mode="valid")
ax.plot(steps[window - 1:], smoothed, color="#27ae60", linewidth=2.5, label="Smoothed")
ax.set_xlabel("Step")
ax.set_ylabel("Loss")
ax.set_title("Training Loss Curve — XTTS-v2 Fine-tuning on Egyptian Arabic", fontsize=13)
ax.grid(True, alpha=0.2)
if len(losses) > 20:
ax.legend()
fig.tight_layout()
fig.savefig(IMAGES_DIR / "training_loss.png", dpi=150)
plt.close(fig)
print(" Saved training_loss.png")
def main():
print("=" * 70)
print(" Phase 7: Evaluation & Comparison")
print("=" * 70)
IMAGES_DIR.mkdir(parents=True, exist_ok=True)
BENCHMARKS_DIR.mkdir(parents=True, exist_ok=True)
# Load audio files
print("\n[1/5] Loading audio files...")
audio_data = {}
sr_data = {}
metrics = {}
for name, path in FILES.items():
if not path.exists():
print(f" WARNING: {path} not found, skipping")
continue
wav, sr = sf.read(str(path))
audio_data[name] = wav
sr_data[name] = sr
m = measure_audio(wav, sr)
metrics[name] = m
print(f" {name}: {m['duration_s']}s, SNR={m['snr_db']}dB, silence={m['silence_ratio']:.1%}")
# Generate charts
print("\n[2/5] Generating waveform comparison...")
plot_waveforms(audio_data, sr_data)
print("\n[3/5] Generating spectrogram comparison...")
plot_spectrograms(audio_data, sr_data)
print("\n[4/5] Generating metrics comparison...")
plot_metrics_comparison(metrics)
print("\n[5/5] Generating training loss curve...")
plot_training_loss()
# Save evaluation report
report = {
"date": __import__("time").strftime("%Y-%m-%d"),
"files": {name: str(path) for name, path in FILES.items()},
"metrics": metrics,
}
report_path = BENCHMARKS_DIR / "evaluation.json"
with open(report_path, "w", encoding="utf-8") as f:
json.dump(report, f, ensure_ascii=False, indent=2)
print(f"\n{'='*70}")
print(" Evaluation Complete!")
print(f"{'='*70}")
print(f"\n Charts saved to: {IMAGES_DIR}/")
print(f" - comparison_waveforms.png")
print(f" - comparison_spectrogram.png")
print(f" - comparison_metrics.png")
print(f" - training_loss.png")
print(f" Report: {report_path}")
if __name__ == "__main__":
main()
|