Extract per-frame audio data.
(path: str, fps: int, n_bands: int)
| 86 | |
| 87 | |
| 88 | def extract(path: str, fps: int, n_bands: int) -> dict: |
| 89 | """Extract per-frame audio data.""" |
| 90 | print(f"Decoding audio from {path}...", file=sys.stderr) |
| 91 | samples = decode_audio(path) |
| 92 | duration = len(samples) / SAMPLE_RATE |
| 93 | frame_step = SAMPLE_RATE // fps |
| 94 | total_frames = int(duration * fps) |
| 95 | |
| 96 | print(f"Duration: {duration:.1f}s, {total_frames} frames at {fps}fps", file=sys.stderr) |
| 97 | print(f"FFT window: {FFT_SIZE} samples ({SAMPLE_RATE / FFT_SIZE:.1f} Hz/bin)", file=sys.stderr) |
| 98 | print(f"Frequency range: {MIN_FREQ:.0f}-{MAX_FREQ:.0f} Hz, {n_bands} bands", file=sys.stderr) |
| 99 | |
| 100 | # Precompute constants |
| 101 | hann = np.hanning(FFT_SIZE) |
| 102 | band_edges = compute_band_edges(n_bands) |
| 103 | freq_per_bin = SAMPLE_RATE / FFT_SIZE |
| 104 | n_bins = FFT_SIZE // 2 + 1 |
| 105 | half_fft = FFT_SIZE // 2 |
| 106 | |
| 107 | # Pass 1: extract raw values |
| 108 | rms_values = np.zeros(total_frames) |
| 109 | band_values = np.zeros((total_frames, n_bands)) |
| 110 | |
| 111 | for f in range(total_frames): |
| 112 | # RMS from the frame's audio slice |
| 113 | rms_start = f * frame_step |
| 114 | rms_end = rms_start + frame_step |
| 115 | frame_slice = samples[rms_start:min(rms_end, len(samples))] |
| 116 | if len(frame_slice) > 0: |
| 117 | rms_values[f] = np.sqrt(np.mean(frame_slice ** 2)) |
| 118 | |
| 119 | # FFT from a centered 4096-sample window |
| 120 | center = rms_start + frame_step // 2 |
| 121 | win_start = center - half_fft |
| 122 | win_end = center + half_fft |
| 123 | |
| 124 | if win_start >= 0 and win_end <= len(samples): |
| 125 | window = samples[win_start:win_end] * hann |
| 126 | else: |
| 127 | # Zero-pad at edges |
| 128 | padded = np.zeros(FFT_SIZE) |
| 129 | src_start = max(0, win_start) |
| 130 | src_end = min(len(samples), win_end) |
| 131 | dst_start = src_start - win_start |
| 132 | dst_end = dst_start + (src_end - src_start) |
| 133 | padded[dst_start:dst_end] = samples[src_start:src_end] |
| 134 | window = padded * hann |
| 135 | |
| 136 | band_values[f] = compute_fft_bands(window, freq_per_bin, n_bins, band_edges, n_bands) |
| 137 | |
| 138 | # Pass 2: normalize |
| 139 | peak_rms = rms_values.max() if total_frames > 0 else 1.0 |
| 140 | if peak_rms > 0: |
| 141 | rms_values /= peak_rms |
| 142 | |
| 143 | # Per-band normalization so treble is visible alongside louder bass |
| 144 | band_peaks = band_values.max(axis=0) |
| 145 | band_peaks[band_peaks == 0] = 1.0 |
no test coverage detected