Files
JustOneCacophony/data/README.en.md
T
TheM14 07a26b41fe Keep only the Rosella and .jochrtf binaural paths
The runtime renders two inputs now: the Rosella .personalized_headphone model and
an already compiled .jochrtf field. SOFA import, the full-HRIR direct renderer
and the SOFA -> .jochrtf fitting compiler leave this checkout; they continue as
an offline tool that only produces caches this repository loads.

The .jochrtf renderer goes back to the implementation that predates the direct
SOFA work: 64-QMF analysis, 77 hybrid bands, per-object hybrid path delays and
the project room (six first-order images plus the shared FDN tail) computed in
code. It is restored verbatim under the .jochrtf vocabulary
(SofaBinauralBackend -> JochrtfBackend) and gains one switch:
--jochrtf-room on|off, off by default. With the room off the renderer outputs
only the field's direct binaural sound, and in either mode it reads no audio
assets.

Because the room is computed in code again, the generated sofa-room*.npz assets
and the modules only the SOFA path used are removed (sofa_canonical, sofa_hrir,
sofa_direct_backend, sofa_room, hybrid_paths, partitioned_convolution,
point_source, compile_hrtf, generate_sofa_room). data/ keeps the decode tables
and the Rosella filterbanks.

The CLI drops --sofa-hrtf, --hrtf-radius-m, --sofa-interpolation,
--hrtf-cache-policy, --hrtf-cache-dir, --sofa-room, --sofa-room-preset and
--source-reference-distance-m. A .jochrtf is still explicit through
--compiled-hrtf-cache, and the default binaural input is the personalized model.

Rosella parsing, filterbank, room and timing are untouched and its output stays
bit-identical. The .jochrtf path reproduces the pre-SOFA renderer sample for
sample, with the room on and with it off.
2026-11-01 19:00:40 +08:00

101 lines
4.2 KiB
Markdown

# Python runtime tables
[中文](README.md)
This directory contains static production tables. It does not contain user HRTFs.
`tables.npz` contains the JOC core decoding tables:
```text
analysis_window float64[10,64]
qmf5_window float64[640]
joc_huff_code_coarse_generic int64[95,2]
joc_huff_code_fine_generic int64[191,2]
joc_huff_code_coarse_coeff_sparse int64[95,2]
joc_huff_code_fine_coeff_sparse int64[191,2]
joc_huff_code_5ch_pos_index_sparse int64[4,2]
joc_huff_code_7ch_pos_index_sparse int64[6,2]
```
`src/joc_qmf.py` loads the QMF tables, while `src/joc_decode.py` loads the JOC Huffman trees.
## Binaural rendering tables
`rosella_kernels.npz` contains the fixed 64-QMF/77-hybrid tables used by
`.jochrtf` rendering and personalized bank 1. Full SOFA rendering uses PCM convolution:
```text
format_version little-endian int32[1]
qmf_analysis_coefficients float32[64,10]
hybrid_analysis_low_kernel float32[3,2,13,16,2]
hybrid_synthesis_indices int16[154,4]
hybrid_synthesis_values float32[154]
qmf_synthesis_basis float64[64,4,128]
qmf_synthesis_taps float64[64,10,4]
```
The float32 table values are promoted to float64 when loaded.
`src/public_filterbank.py` verifies the archive and every array by SHA-256.
Those hashes, the table version, and the 77 reference band-center values all
participate in the `.jochrtf` cache key. The full analysis/synthesis latency is
961 samples.
The packaged tables implement publicly standardized filter banks, computable
from the following formulas.
The 64-QMF → 77-hybrid structure, the 13-tap low-band prototypes, and their
half-bin complex modulation are defined in
[3GPP TS 26.405 / ETSI TS 126 405](https://www.etsi.org/deliver/etsi_ts/126400_126499/126405/06.00.00_60/ts_126405v060000p.pdf),
Section 5.2.2 (Table 1 $Q=8$/$Q=4$ coefficients, delay 6):
$$G_q^p[n] = g^p[n]\cdot\exp\!\Bigl(j\,\frac{2\pi}{Q^p}\bigl(q+\tfrac12\bigr)(n-6)\Bigr),\qquad n=0,\dots,12$$
The 64-band QMF analysis is the MPEG-4 AAC/SBR 64 complex QMF analysis bank of
ISO/IEC 14496-3/AMD1:2003, subclause 4.B.18.2; the packaged $64\times10$ table
is the polyphase reordering of the public 640-tap prototype $c_0,\dots,c_{639}$:
$$A_{r,t} = \frac{(-1)^t}{128}\,c_{63-r+64t},\qquad r=0,\dots,63,\ t=0,\dots,9$$
The QMF synthesis table is the causal left inverse of the analysis polyphase
matrix $\mathbf{A}$, i.e. the solution of $\mathbf{A}\,\mathbf{W}=\mathbf{P}$
($\mathbf{P}$ is the 577-sample delay permutation; total latency
$961 = 577 + 6\times64$), stored as a rank-4 factorization:
$$W_{b,l} = \sum_{r=1}^{4} t_{b,l,r}\,\mathbf{b}_{b,r}^{\top}$$
The hybrid synthesis table is the 77→64 recombination: identity for the high
bands, $Y_{3+b}=X_{16+b}$, and for the low bands ($C_p$ is the $8+4+4$ child
partition):
$$Y_p = \sum_{q\in C_p}\Bigl(\operatorname{Re}X_q + j\,s_q\,\operatorname{Im}X_q\Bigr),\qquad s_q\in\{\pm1\}$$
The same values also appear in other public implementations of these standards
(for example FFmpeg's `aacps_tablegen.h` and `aacsbrdata.h`).
### Captured Rosella filterbank
`rosella_captured_kernels.npz` holds the bank the reference headphone renderer
actually runs (`--rosella-filterbank 2`):
```text
format_version little-endian int32[1]
qmf_prototype float32[64,10]
rotation complex64[64]
hybrid complex64[22]
```
The table values are promoted to float64/complex128 on load. The full
analysis/synthesis latency is 448 samples (961 for the public bank). These
values come from runtime buffer captures of the reference renderer, not from the
public standards above; the C++ core compiles the same data in as
`src/hrtf/rosella_filterbank_tables.h`. This file also contains no personalized
HRTF or room tuning.
Public availability of a standard does not by itself grant permission to
practice related patent claims.
SOFA is the user-visible source of truth. A `.jochrtf` file is a disposable JOC
compiled HRTF cache that can be rebuilt from SOFA. The cache contains
transformed source-HRTF data and remains subject to the source SOFA/HRTF
dataset's licence and redistribution restrictions.