One full-decode control and one adjustable windowed player per codec. Move a slider to hear what changes, without losing your place in the clip.
Highlighted points are the windows selected above. Click a point to move that codec’s slider.
Original: the complete source after mono downmix and 24 kHz metric-domain resampling. Full decode: its cached code tensor decoded in one call. Window decode: the exact same tensor split into independent windows, decoded without shared state or overlap, then hard-concatenated. All codebooks are retained. The last short window is kept, and outputs are cropped to source length.
All players use 24 kHz mono IEEE float32 WAV: no lossy export compression, loudness normalization or integer quantization. Source compression and volume differences are retained. A browser or output device may clip peaks above full scale; the manifest records each asset’s peak.
One source per dataset is held fixed across all codecs, metrics and bands. Clips were selected from the existing 60-clip study for complete band coverage, then proximity to the dataset’s median duration. In error-band mode, each width is closest to its band midpoint; the >50% example targets 75%. Window-size mode includes every distinct exported window across all three metric ladders; its slider is discrete, not interpolated, and does not include unexported sweep conditions. These are illustrative examples, not a representative listening trial or a new music/multilingual evaluation.
Band formula: 100 × (clip source-versus-window error / clip source-versus-full error − 1). The first band includes 0 and 5%; subsequent bands exclude the lower bound and include the upper. The final band is strictly above 50%. Window-size mode may expose negative-error cases for the selected metric. Different metrics may select different widths for the same error band.
Cached codes, source hashes and pinned checkpoints were checked. Full-control errors agree with recorded measurements within 10⁻⁶; treatment errors within 10⁻⁵. Exact errors, source IDs, discrete geometry and audio hashes are in the manifest. The original measurements and thresholds are unchanged.