Render SOFA directly with full HRIR interpolation

SOFA inputs now use delay-aligned local interpolation and partitioned
convolution instead of compiling a compressed directional field.

- preserve full FIR support, complex phase and explicit delay handling
- add log-magnitude/phase and local magnitude-corrected interpolation
- support Near/Mid/Far with a shared calibration anchor
- retain .jochrtf as an explicit lightweight input and offline export
- keep personalized headphone rendering unchanged
- update the CLI, reports and bilingual documentation

Checked: 21 automated tests, 8 bit-exact personalized regressions,
and full-song D2 renders in Near/Mid/Far.
This commit is contained in:
TheM14 committed 2026-11-01 00:57:49 +08:00
1 parent 0e98e91762
commit faf8a921c2
13 files changed
+1140 -591

No files matched your search

+36 -253
View File
@@ -2,283 +2,66 @@
[中文](binaural.md) · [Back to README](../README.en.md)
JustOneCacophony's binaural backend supports three HRTF sources:
`SimpleFreeFieldHRIR` SOFA, the Rosella `.personalized_headphone` model exported
by Dolby's official personalization scan (its JSON parsing is implemented by
this project and invokes no Dolby software), and the `.jochrtf` cache compiled
from SOFA. SOFA is compiled into an in-memory directional field when the model
is loaded. A `.jochrtf` file is only a disposable, reproducible JOC compiled
HRTF cache; it is neither an interchange format nor a prerequisite for using
SOFA.
## Routes
```text
SOFA FIR
-> CanonicalHrtf
-> 48 kHz / one radius shell / delay-phase policy
-> 64-QMF / 77-hybrid projection
-> fifth-order ACN/N3D real-SH field
-> per-object direct + early reflections
-> shared unitary-FDN late room
-> float64 stereo
```
- SOFA: full complex local HRIR interpolation, adaptive-length filters and 128-sample partitioned convolution.
- `.jochrtf`: the existing compact 77-band/fifth-order field and delay representation.
- `.personalized_headphone`: the existing DSP, filterbank selection and Near/Mid/Far behavior.
## Inputs
The CLI has three mutually exclusive HRTF input sources; with none given, a
default rule resolves the input:
SOFA and compact inputs use the same externally configured reference geometry, distance gains and environment state. SOFA never compiles or silently falls back to the compact format. The old compact loader/writer and compiler remain available.
```powershell
# 1) SOFA: defaults to HRTF/binaural.sofa, or an explicit path
python main.py input.m4a --binaural
python main.py input.m4a --binaural --sofa-hrtf C:\HRTF\subject.sofa
# 2) Rosella .personalized_headphone: defaults to HRTF/binaural.personalized_headphone
python main.py input.m4a --binaural --personalized-headphone
python main.py input.m4a --binaural --personalized-headphone C:\HRTF\subject.personalized_headphone
# 3) .jochrtf: explicitly load a compiled cache
python main.py input.m4a --binaural `
--compiled-hrtf-cache C:\HRTF\subject.jochrtf
# Optional: create/reuse a transparent disk cache for SOFA
python main.py input.m4a --binaural --sofa-hrtf C:\HRTF\subject.sofa `
--hrtf-cache-policy disk
python main.py input.m4a --binaural --sofa-hrtf subject.sofa --reference-config reference.json
python main.py input.m4a --binaural --sofa-hrtf subject.sofa --sofa-interpolation lf_mca_zero
python main.py input.m4a --binaural --sofa-hrtf subject.sofa --binaural-mode near
python main.py input.m4a --binaural --sofa-hrtf subject.sofa --binaural-mode far
python src\compile_hrtf.py subject.sofa subject.jochrtf
python main.py input.m4a --binaural --compiled-hrtf-cache subject.jochrtf
python main.py input.m4a --binaural --personalized-headphone --binaural-mode far --rosella-filterbank 2
```
The default order is `HRTF/binaural.sofa`, then the unique `.jochrtf` under
`output/hrtf-cache`, then `HRTF/binaural.personalized_headphone`; if none of
the three exist, an error asks for an explicit path. Multiple `.jochrtf` files
under `output/hrtf-cache` are also an error requiring an explicit choice.
The default SOFA is `HRTF/binaural.sofa` under the project directory, independently of the current working directory. Use `--sofa-hrtf PATH` for other locations. A single measured shell is selected automatically; multi-shell data requires `--hrtf-radius-m`. The personalized default path is unchanged. All three binaural routes accept `--binaural-mode near|mid|far`, defaulting to mid in the CLI; off remains an ADM-only DBMD hint.
The `.personalized_headphone` JSON parsing is implemented by this project
(`src/rosella_model.py`) and does not invoke any Dolby software.
## Reference configuration
`--hrtf-cache-policy` accepts `none`, `memory`, or `disk`. The default is
`memory`; neither `none` nor `memory` creates a file. `disk` writes to
`output/hrtf-cache` by default, or to `--hrtf-cache-dir`. `--hrtf-radius-m`
selects the nearest measurement-radius shell.
Resolution order: `--reference-config`, `BINAURAL_REFERENCE_CONFIG`, then local `local/binaural_reference.json`. Missing data, a hash mismatch, or an unsupported mapping is an error. Environment data and its specific identity are not supplied with the source. Example:
The Python API also uses explicit factories:
```python
from sofa_binaural_backend import SofaBinauralBackend
renderer = SofaBinauralBackend.from_sofa(
"subject.sofa",
source_count=16,
default_profile="mid",
cache_policy="memory",
)
cached = SofaBinauralBackend.from_compiled_cache(
"subject.jochrtf",
source_count=16,
default_profile="mid",
)
```json
{"schema":"binaural-reference-v1","profile_id":"reference-v1",
"model_path":"C:/path/to/reference-data.json","model_sha256":"<64 hex digits>",
"profile_index":3,"room_impulse_slots":4096,"level_calibration_directions":128,
"compiled_receiver_positions_m":[[0,0.09,0],[0,-0.09,0]]}
```
The factories never guess a format from an unknown suffix: SOFA and `.jochrtf`
always use distinct loaders.
The external data must be accepted by the existing parameter loader. Relative paths resolve against the configuration file. `profile_index` fixes the shared level-calibration anchor; `--binaural-mode` selects the rendering profile without changing that anchor. Compact files do not store receiver positions; they require explicit configured geometry. Full SOFA uses its own ear positions. A measured radius does not set the scene scale or room size.
## Binaural render mode
Reference mapping uses the selected profile's quantization, axes, bounds and centre regularization. Outside positions extend the listener-to-boundary ray by the coded finite factor and are never clamped back inside. Near uses direct gain 1 and environment send 0. Mid/Far use `a=1/sqrt(1+c*(h0*r)^2)` and `b=h0*r*a`. Send gain is not final reverberant pressure. All modes share the same interpolation, radial adaptation, convolution and timing. There is no additional 1/r attenuation or absolute flight-time delay.
`--binaural-mode off|near|mid|far` (default `mid`) is a **human-specified
rendering hint**, not original binaural metadata extracted or recovered from the
input E-AC-3 JOC bitstream:
## Full HRIR model
- Direct binaural rendering (`--binaural`): near/mid/far apply, default `mid`;
`off` is an error;
- ADM BWF: the low 3 binaural-render-mode bits of the last 15 JOC object entries
in DBMD segment 10 carry `off=0/near=1/far=2/mid=3`, leaving the first 10 bed
entries unchanged; the default is `mid`, and `off` explicitly disables the
binaural metadata hint.
The strict importer validates SimpleFreeFieldHRIR 0.4/1.0/1.1, FIR, units, listener orientation, ear order, zero emitter offset, finite data and free-field declarations. Resampling scales Data.Delay in samples. The radial adapter requires fixed symmetric receiver positions centred on the listener.
## Canonical SOFA contract
Map ADM right/front/up to SOFA front/left/up with `(y,-x,z)`. A 3D convex hull supplies local spherical triangles. Ray/planar barycentric weights are nonnegative and sum to one; exact measured nodes preserve the full complex response. Large triangles over coverage holes still have limited support.
The strict importer currently accepts:
Per-ear delay comes from phase regression at 200–1500 Hz. Interpolate aligned complex residuals and restore weighted delay, including Data.Delay once. The default `lf_logmag_phase` combines interpolated log magnitude with aligned complex phase. `lf_mca_zero` is an independent local MCA approximation with 41 continuous Gammatone power bands, PCHIP and a ±12 dB gain bound; it is not a bitwise reproduction of a third-party implementation.
- `Conventions=SOFA`;
- `SOFAConventions=SimpleFreeFieldHRIR`, version `0.4`, `1.0`, or `1.1`;
- `DataType=FIR` and `Data.IR[M,2,N]`;
- one positive finite `Data.SamplingRate` in hertz/Hz;
- spherical or Cartesian `SourcePosition`;
- singleton or per-measurement `ListenerPosition/View/Up`;
- two receivers whose listener-local lateral geometry uniquely identifies L/R;
- one zero-offset emitter;
- causal `Data.Delay[I,2]` or `[M,2]`;
- an explicitly free-field/anechoic `RoomType`.
Radial adaptation multiplies by the complex normalized rigid-sphere ratio `H_sphere(r)/H_sphere(r_measured)`, exactly one at the measured radius. Ear geometry determines sphere radius. Near the centre, the configured regularization and an equal-ear, zero-ITD, spherical-area-weighted energy response give a continuous blend. This is an approximate human-head radial model; angular interpolation scores do not validate arbitrary distances.
Receiver order comes from geometry, never from the receiver array index. SOFA
listener coordinates are $+X$ front,
$+Y$ left,
$+Z$ up; ADM coordinates are
$+X$ right,
$+Y$ front,
$+Z$ up:
Calibration computes one global scalar per input at the configured anchor from equal-area directions and total binaural energy with logarithmic frequency weighting at 100–16000 Hz. All three modes share that scalar; no per-mode/ear/direction/frequency normalization is applied.
$$\bigl(x_{\mathrm{SOFA}},\ y_{\mathrm{SOFA}},\ z_{\mathrm{SOFA}}\bigr) = \bigl(y_{\mathrm{ADM}},\ -x_{\mathrm{ADM}},\ z_{\mathrm{ADM}}\bigr)$$
The FFT size is at least 2048, automatically increased to a power of two that fits the complete HRIR and Data.Delay, up to 16384. All resulting FIR taps are retained, with a declared common shift of half the FFT size. Rendering rebases one global recording origin, the rounded median of mean-ear fitted delays; the environment branch receives matching delay before common compensation. Original spectrum queries retain the dataset time reference. The shift is independent of target source distance.
`CanonicalHrtf` keeps `Data.IR` and `Data.Delay` separate. Only a time-domain
baseline calls `materialized_measurement()` to apply delay once; the runtime SH
path never materializes and then restores the delay. Non-48-kHz HRIRs are
normalized with float64 `scipy.signal.resample_poly`, and delay samples scale by
the same ratio.
Per-object convolution retains input history and independent old/new overlap during a 128-sample output crossfade. Environment computation batches 1536 samples without changing the 128-sample parameter timeline; extra computation padding is trimmed when draining. Source PCM, delays and accumulation use float64; spectra use complex128, at 48 kHz.
GeneralFIR, BRIR, TF, multiple emitters, ambiguous receivers, and non-free-field
data require convention-specific adapters. They cannot enter the core importer
through a reshape.
## Scheduling and limits
## Exactly-once delay and phase
Each `[1536,16]` frame contains LFE followed by 15 already reconstructed objects/bed slots. Metadata retains the existing 1473-sample content offset and 32-sample alignment; new parameters update every 128 samples. A SOFA-only opt-in preserves negative elevation, outside distance and anchor fields. Legacy parsing and the personalized timeline retain their default behavior.
The compiler recognizes three mutually exclusive representations:
Multi-position blocks, ISF, explicit infinity and screen anchors requiring calibration return structured unsupported errors. Finite outside updates participate in timeline interpolation. `finish()` drains state only once; output gain is applied once without an implicit limiter.
1. Nonzero `Data.Delay` is external to `Data.IR`; the FIR is not de-rotated and
runtime applies the delay once.
2. With `Data.Delay=0` and an ordinary positive-onset HRIR, each ear's main peak
supplies arrival time. Compilation separates it and runtime restores it once.
The current threshold is a peak index greater than two samples.
3. With `Data.Delay=0` and both FIRs at a shared sample-zero origin, no external
delay is invented. The authored complex phase stays in the fifth-order field.
## Compact files
No path may add a second ear delay or phase-group delay.
Version-1 `.jochrtf` files retain their pickle-free ZIP schema, band centres, complex128 `[36,2,77]` coefficients, float64 `[36,2]` delay coefficients, `[2,2]` bounds, metadata, hash validation and atomic writer. Their inherent directional compression error remains. See [filterbank data](../data/README.en.md) for the public filterbank sources.
## Public filterbank and directional field
The runtime is fixed at:
- 48 kHz;
- a 64-sample QMF hop;
- 64-QMF / 77 hybrid bands;
- 961 samples of analysis/synthesis latency;
- fifth order, 36 terms, ACN/N3D real spherical harmonics;
- float64 PCM, delay, SH, and room state; complex128 band transfers and spectra.
Those are the SOFA / `.jochrtf` path and the public filterbank. The Rosella
`.personalized_headphone` path selects its bank: the public bank above is the
default (961 samples), and `--rosella-filterbank 2` switches to the bank captured
from the reference renderer (11 taps / 3 slots, 448 samples). Parameters,
buffers, latency and validation of both are in
[Rosella filterbank](rosella_filterbank.en.md).
Real and imaginary unit gains for every hybrid band pass through the same
analysis/synthesis chain to form a 154-real-parameter impulse dictionary. The
compiler does not sample 77 FFT bins. Defaults are `1e-3` projection ridge and
`1e-5` SH ridge. Coincident directions are merged before a spherical-Voronoi
weighted ridge fit.
The fixed resource is `data/rosella_kernels.npz`, which implements publicly
standardized filter banks, computable from the following formulas.
The hybrid analysis kernels 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 QMF analysis table is the MPEG-4 AAC/SBR 64 complex QMF bank of
ISO/IEC 14496-3/AMD1:2003, subclause 4.B.18.2, stored as 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):
$$Y_p = \sum_{q\in C_p}\Bigl(\mathrm{Re}X_q + j\,s_q\,\mathrm{Im}X_q\Bigr),\qquad s_q\in\{\pm1\}$$
The loader verifies the archive and every array by SHA-256; the table version,
all array hashes, and the 77 reference band-center values are part of the cache
key. Public availability of a standard does not by itself grant permission to
practice related patent claims. See
[`data/README.en.md`](../data/README.en.md) and
[`THIRD_PARTY_NOTICES.md`](../THIRD_PARTY_NOTICES.md) for the sources and the
rights boundary.
## `.jochrtf`
A `.jochrtf` file is a pickle-free compressed NumPy archive with an exact member set:
| key | dtype / shape |
|---|---|
| `metadata_json` | NumPy Unicode scalar containing JSON text (`dtype.kind == "U"`) |
| `band_center_frequencies_hz` | little-endian `float64[77]` |
| `coefficients` | little-endian `complex128[36,2,77]` |
| `delay_coefficients` | little-endian `float64[36,2]` |
| `delay_bounds` | little-endian `float64[2,2]` |
Metadata uses the `JOC-HRTF-CACHE` magic and records the schema, compiler and
phase-policy versions, ACN/N3D convention, filterbank hashes, SOFA content
SHA-256, sample rate, radius, order, both ridge values, payload hash, and fit
report. Every setting that changes compilation participates in the cache key.
Metadata never persists an absolute local `source_path`; it may keep a display
name only.
Before constructing a field, the loader uses `allow_pickle=False` and validates
ZIP members and expanded sizes, shapes, dtypes, byte order, contiguous layout,
finite values, delay bounds, band centers, payload hash, and cache key. The
writer uses a same-directory temporary file, `fsync`, a process-held OS file
lock, and atomic `os.replace`. Its hidden `.lock` sidecar may remain and does not
mean that a writer still owns the lock. Outdated, damaged, or mismatched
caches cannot hit. SOFA input rebuilds an invalid cache; an explicitly selected
cache reports the error.
Deleting a disk cache must not change the field or render produced from the same
SOFA and compiler configuration.
A `.jochrtf` file contains directional-field coefficients and delay data
transformed from the source HRIRs. Its reproducibility therefore does not make
it licence-free. Creating a cache does not enlarge the rights granted by the
source SOFA/HRTF dataset: use, copying, and redistribution remain subject to
that dataset's terms. If those terms are unclear, keep `.jochrtf` as a private
local cache and do not ship it with the program or another build artifact.
`source_sha256` is only a content-integrity identifier, not proof of provenance
or permission.
## JOC objects and room behavior
The production adapter retains the existing JOC schedule:
- `[1536,16]` input per frame;
- channel 0 is special LFE and channels 1..15 are JOC objects;
- ID11/OAMD positions use a sample-timed timeline;
- source parameters update every 512 samples;
- every object owns independent direct/early history while one late FDN is shared;
- `finish()` drains early/late tails; output gain is explicit, with no implicit
limiter or programme loudness normalization.
Near/Mid/Far, equal-power direct level, six first-order shoebox image sources,
late sends, the unitary FDN, the 120–180 Hz cosine-squared LFE low-pass, and room
calibration are JOC project-defined behavior, not constants published by SOFA or
Dolby.
The filterbank, the SH direction-field evaluation, the per-object direct/early
histories and the shared FDN all run in Python/NumPy, with the binaural DSP kept
in float64/complex128 throughout.
## Technical references and rights boundary
- [SOFA SimpleFreeFieldHRIR convention](https://www.sofaconventions.org/mediawiki/index.php/SimpleFreeFieldHRIR)
- [3GPP TS 26.405 / ETSI TS 126 405 (64-QMF/77-hybrid definition)](https://www.etsi.org/deliver/etsi_ts/126400_126499/126405/06.00.00_60/ts_126405v060000p.pdf)
- [Dolby binaural render-mode workflow](https://professionalsupport.dolby.com/s/article/What-is-Binaural-Render-Mode-and-how-do-the-settings-affect-my-mix)
- [EP3090576A1](https://patents.google.com/patent/EP3090576A1/en), used only as
architectural background for direct/early/late, subbands, and FDNs; it does
not establish that any product uses a particular embodiment.
Public availability of a specification, source file, or patent document does
not by itself authorize copying its contents, redistribution of derivatives,
or practice of patent claims. These technical references grant no patent
licence and make no non-infringement representation. Anyone preparing a release
or product integration must assess the applicable data and software licences,
patent permissions, and freedom to operate. See
[`THIRD_PARTY_NOTICES.md`](../THIRD_PARTY_NOTICES.md) for the public-standard
provenance and rights boundary.
Mathematical sources: [Duda and Martens 1998](https://www.ece.ucdavis.edu/cipic/wp-content/uploads/sites/12/2015/04/cipic_JASA_Nov_1998.pdf), [MCA paper](https://doi.org/10.1109/TASLP.2023.3313908), [ETSI TS 103 420](https://www.etsi.org/deliver/etsi_ts/103400_103499/103420/01.02.01_60/ts_103420v010201p.pdf). HRTF algorithms are independently implemented; external data retain their own licence terms.