"""Mini self-contained verification test for open-xtts-trainer. Generates synthetic audio tones and tests Stage 1 (DVAE), Stage 2 (Vocab Extender), and Stage 3 (HiFi-GAN) forward-backward loops on CPU in under 15 seconds. """ import os import shutil import tempfile import torch import torchaudio from xtts_trainer.config import AudioConfig, HiFiGANConfig, DVAEConfig from xtts_trainer.stage1_dvae.train import train_dvae from xtts_trainer.stage2_gpt.vocab_extender import extend_vocab from xtts_trainer.stage3_hifigan.extract_latents import extract_latents from xtts_trainer.stage3_hifigan.train import train_hifigan def generate_synthetic_audio(dir_path: str, num_files: int = 4, duration_sec: float = 1.0, sr: int = 24000): """Generates short synthetic sine waves for testing.""" import numpy as np from scipy.io import wavfile os.makedirs(dir_path, exist_ok=True) num_samples = int(duration_sec * sr) t = np.linspace(0, duration_sec, num_samples, endpoint=False) for i in range(num_files): freq = 220.0 * (i + 1) audio = (0.5 * np.sin(2 * np.pi * freq * t)).astype(np.float32) file_path = os.path.join(dir_path, f"sample_{i+1:03d}.wav") wavfile.write(file_path, sr, audio) print(f"Generated {num_files} synthetic audio files in: {dir_path}") def test_stage2_vocab_extender(tmp_dir: str): print("\n--- Testing Stage 2: Vocab Extender ---") base_vocab_path = os.path.join(tmp_dir, "vocab.json") target_corpus_path = os.path.join(tmp_dir, "target.txt") out_vocab_path = os.path.join(tmp_dir, "vocab_extended.json") # Base vocab with 5 tokens import json base_vocab = {"[PAD]": 0, "[UNK]": 1, "a": 2, "b": 3, "c": 4} with open(base_vocab_path, "w") as f: json.dump(base_vocab, f) # Target text with Arabic / Hindi characters with open(target_corpus_path, "w", encoding="utf-8") as f: f.write("مرحبا नमस्ते") new_vocab = extend_vocab(base_vocab_path, target_corpus_path, out_vocab_path) assert len(new_vocab) > len(base_vocab), "Vocab was not extended!" print("Stage 2 Test Passed!") def test_stage1_dvae(tmp_dir: str, audio_dir: str): print("\n--- Testing Stage 1: DVAE Training ---") dvae_out = os.path.join(tmp_dir, "dvae_ckpt") cfg = DVAEConfig() cfg.batch_size = 2 cfg.num_tokens = 512 cfg.hidden_dim = 64 cfg.codebook_dim = 64 # Run for 2 epochs on CPU train_dvae( audio_dir=audio_dir, output_dir=dvae_out, config=cfg, epochs=2, batch_size=2, device="cpu", ) assert os.path.exists(os.path.join(dvae_out, "dvae_epoch_2.pth")), "DVAE checkpoint not saved!" print("Stage 1 Test Passed!") def test_stage3_hifigan(tmp_dir: str, audio_dir: str): print("\n--- Testing Stage 3: Latent HiFi-GAN Vocoder ---") latents_dir = os.path.join(tmp_dir, "latents") hifigan_out = os.path.join(tmp_dir, "hifigan_ckpt") # 1. Extract latents extract_latents(audio_dir, latents_dir, device="cpu") # 2. Train HiFi-GAN cfg = HiFiGANConfig() cfg.batch_size = 2 cfg.segment_size = 4096 # small segment for fast CPU test train_hifigan( audio_dir=audio_dir, latents_dir=latents_dir, output_dir=hifigan_out, config=cfg, epochs=2, batch_size=2, device="cpu", ) assert os.path.exists(os.path.join(hifigan_out, "hifigan_epoch_2.pt")), "HiFi-GAN checkpoint not saved!" print("Stage 3 Test Passed!") def main(): print("==================================================") print("Running open-xtts-trainer Sanity Test Suite (CPU)") print("==================================================") tmp_dir = tempfile.mkdtemp(prefix="open_xtts_test_") audio_dir = os.path.join(tmp_dir, "audio") try: generate_synthetic_audio(audio_dir, num_files=4, duration_sec=0.5) test_stage2_vocab_extender(tmp_dir) test_stage1_dvae(tmp_dir, audio_dir) test_stage3_hifigan(tmp_dir, audio_dir) print("\n ALL 3 STAGES PASSED THE SANITY TEST!") finally: shutil.rmtree(tmp_dir, ignore_errors=True) if __name__ == "__main__": main()