import numpy as np import pytest import soundfile as sf from app.audio import ( AudioBuffer, AudioError, analyze, analyze_file, hpss_separate, hpss_separate_file, load_audio, mixture_report, save_stems, ) def synth(n=12000, sr=12000): t = np.arange(n) / sr signal = 0.15 * np.sin(2 * np.pi * 330 * t) for p in (1500, 4500, 7800): pulse = np.exp(-np.arange(70)/15) signal[p:p+len(pulse)] += 0.5 * pulse return AudioBuffer(np.stack((signal, signal * 0.8), axis=1), sr) def test_real_hpss_reconstructs_stereo(tmp_path): audio = synth() stems = hpss_separate(audio) assert list(stems) == ["harmonic", "percussive", "residual"] assert all(v.shape == audio.samples.shape for v in stems.values()) report = mixture_report(audio.samples, stems) assert report["consistent"] assert report["max_abs_reconstruction_error"] < 1e-12 assert np.max(np.abs(stems["harmonic"])) > 1e-4 assert np.max(np.abs(stems["percussive"])) > 1e-4 saved = save_stems(tmp_path, audio, stems) assert set(saved) == set(stems) for stem in saved.values(): x, rate = sf.read(tmp_path / stem["master"], always_2d=True) assert x.shape == audio.samples.shape and rate == audio.samplerate assert (tmp_path / stem["preview"]).exists() def test_silence_supported(): audio = AudioBuffer(np.zeros((1024, 1)), 8000) stems = hpss_separate(audio) assert mixture_report(audio.samples, stems)["consistent"] assert analyze(audio)["peak_dbfs"] == -240.0 def test_corrupted_input_rejected(tmp_path): source = tmp_path / "fake.wav" source.write_text("not audio") with pytest.raises(AudioError): load_audio(source) def test_audio_over_180_seconds_has_no_duration_ceiling(tmp_path): source = tmp_path / "long_recording.wav" with sf.SoundFile(source, "w", samplerate=8000, channels=1) as target: target.seek(8000 * 181 - 1) target.write(np.zeros((1,), dtype=np.float32)) stats = analyze_file(source, tmp_path / "original_preview.wav") assert stats["duration_seconds"] == 181 assert stats["frames"] == 8000 * 181 assert (tmp_path / "original_preview.wav").exists() assert load_audio(source).samples.shape == (8000 * 181, 1) stems, report = hpss_separate_file(source, tmp_path / "stems") assert report["consistent"], report assert all(sf.info(tmp_path / "stems" / info["master"]).frames == 8000 * 181 for info in stems.values()) def test_streamed_hpss_handles_non_silent_stereo_chunk_boundaries(tmp_path): original = synth(n=30000, sr=12000) source = tmp_path / "song.wav" sf.write(source, original.samples, original.samplerate, subtype="FLOAT") stem_dir = tmp_path / "stems" fractions = [] entries, report = hpss_separate_file( source, stem_dir, progress=fractions.append, chunk_frames=8192, margin_frames=4096, ) assert len(fractions) >= 3 and fractions[-1] == 1 assert report["consistent"] exports = [] for item in entries.values(): loaded, sr = sf.read(stem_dir / item["master"], always_2d=True) assert sr == original.samplerate and loaded.shape == original.samples.shape exports.append(loaded) assert np.max(np.abs(sum(exports) - original.samples)) < 1e-6 assert np.max(np.abs(exports[0])) > 0.01 assert np.max(np.abs(exports[1])) > 0.01