Spaces:
				
			
			
	
			
			
					
		Running
		
	
	
	
			
			
	
	
	
	
		
		
					
		Running
		
	| # Copyright (c) Meta Platforms, Inc. and affiliates. | |
| # All rights reserved. | |
| # | |
| # This source code is licensed under the license found in the | |
| # LICENSE file in the root directory of this source tree. | |
| import torch | |
| from audiocraft.models.watermark import AudioSeal | |
| from tests.common_utils.wav_utils import get_white_noise | |
| class TestWatermarkModel: | |
| def test_base(self): | |
| sr = 16_000 | |
| duration = 1.0 | |
| wav = get_white_noise(1, int(sr * duration)).unsqueeze(0) | |
| wm = AudioSeal.get_pretrained(name="base") | |
| secret_message = torch.randint(0, 2, (1, 16), dtype=torch.int32) | |
| watermarked_wav = wm(wav, message=secret_message, sample_rate=sr, alpha=0.8) | |
| result = wm.detect_watermark(watermarked_wav) | |
| detected = ( | |
| torch.count_nonzero(torch.gt(result[:, 1, :], 0.5)) / result.shape[-1] | |
| ) | |
| detect_prob = detected.cpu().item() # type: ignore | |
| assert detect_prob >= 0.0 | |