import tempfile import unittest from pathlib import Path import numpy as np import torch from fiato.config import Config from fiato.export import Exportable, export from fiato.model import Fiato from fiato.text import PHONE_IDS def small_model(): torch.manual_seed(0) model = Fiato(Config()).eval() # A fresh model predicts arbitrary durations; pin them to a few frames per token. torch.nn.init.zeros_(model.duration.out.weight) model.duration.out.bias.data[:] = torch.tensor([1.5, 1.0]) torch.nn.init.normal_(model.vocoder.harmonic.weight, std=0.05) torch.nn.init.normal_(model.vocoder.noise.weight, std=0.05) return model def utterance(length=12): phones = sorted(PHONE_IDS) tokens = torch.tensor([[phones[(7 * i) % len(phones)] for i in range(length)]]) return tokens, (torch.arange(length)[None] % 3) class ExportTest(unittest.TestCase): def test_rewrite_matches_synthesize(self): model = small_model() tokens, stress = utterance() for voice in (0, 1): reference, lengths, details = model.synthesize(tokens, stress, torch.tensor([tokens.shape[1]]), torch.tensor([voice]), speed=1.25, pitch_shift=0.1, pitch_range=1.3) noise = torch.randn(int(lengths[0]), generator=torch.Generator().manual_seed(0))[None] one = lambda value: torch.tensor([value]) with torch.no_grad(): audio, durations = Exportable(model, external_noise=True)(tokens, stress, torch.tensor([voice]), one(1.25), one(0.1), one(1.3), noise) self.assertTrue(torch.equal(durations, details["durations"])) self.assertEqual(audio.shape, reference.shape) self.assertLess(float((audio - reference).abs().max()), 1e-2 * float(reference.abs().max())) def test_onnx_runs_at_other_lengths(self): try: import onnxruntime as ort except ImportError: self.skipTest("onnxruntime is not installed") model = small_model() with tempfile.TemporaryDirectory() as directory: path = export(model, Path(directory) / "model.onnx", external_noise=True) session = ort.InferenceSession(str(path), providers=["CPUExecutionProvider"]) for length in (5, 40): tokens, stress = utterance(length) reference, lengths, _ = model.synthesize(tokens, stress, torch.tensor([length]), torch.tensor([1])) noise = np.zeros((1, model.config.hop * model.config.max_frames), dtype=np.float32) noise[0, :int(lengths[0])] = torch.randn(int(lengths[0]), generator=torch.Generator().manual_seed(0)).numpy() one = lambda value: np.array([value], dtype=np.float32) audio, _ = session.run(None, {"tokens": tokens.numpy(), "stress": stress.numpy(), "voice": np.array([1]), "speed": one(1.0), "pitch_shift": one(0.0), "pitch_range": one(1.0), "noise": noise}) self.assertEqual(audio.shape, tuple(reference.shape)) self.assertLess(float(np.abs(audio - reference.numpy()).max()), 2e-2 * float(reference.abs().max())) if __name__ == "__main__": unittest.main()