Paper Companion Demonstration

An Efficient Parametric Codec for Low-Bitrate First-Order Ambisonics

Wei-Ting Lai, Amy Bastine, Lachlan Birnie, Thushara D. Abhayapala, Prasanga N. Samarasinghe

Audio & Acoustic Signal Processing Group, The Australian National University, Canberra, Australia

For the best 3D spatial audio experience, we recommend wearing headphones.

Listening Direction

Azimuth θ 0.0°
Elevation φ 0.0°
Drag to rotate listening direction

On supported phones and tablets, tap Motion Sensor and rotate your device to change the listening direction. Allow sensor access if prompted.

Audio Comparison

STARSS23 · Full recording
0:00.0 0:44.7
90%

Switch versions while rotating the sound field (keys 1, 2, 3, 4). Bitrates shown are nominal totals for audio and spatial metadata.

Lightweight Spatial Audio Compression, Built to Pair with Any Monaural Codec

MetaFOA takes a lightweight parametric approach to 360° spatial audio. A unified 3-D Directivity Vector combines direction and diffuseness in one representation, while cross-band quantization compresses the spatial metadata to just 0.75 kbps. Pairing this compact representation with a pretrained monaural codec keeps the system modular, efficient, and easy to adapt.

Plug-and-Play with Any Monaural Codec

MetaFOA cleanly separates the omnidirectional audio signal (Wt) from the spatial cues. The audio core can be compressed by any off-the-shelf monaural codec—such as EnCodec, DAC, or future neural audio tokenizers—entirely pretrained and frozen, with zero fine-tuning required.

Unified 3-D Directivity Vector & Fast Fitting

Instead of encoding sound direction (DOA) and room diffuseness separately, MetaFOA merges them into a single 3-D directivity vector and compresses all frequency bands jointly via Cross-Band Residual Vector Quantization (RVQ). The entire codebook fits offline via stage-wise k-means in under 7 minutes without backpropagation.

Figure 1: Overview of the proposed parametric FOA codec architecture showing the frozen monaural Signal Codec branch and the proposed Cross-band RVQ Metadata Codec branch.
Figure 1. Overview of the proposed MetaFOA architecture. The omnidirectional channel Wt is processed by any pretrained monaural codec, while bandwise 3-D directivity vectors vt,1:B are jointly quantized via Cross-Band RVQ at a constant bitrate. PDF ↗

How MetaFOA Preserves Fine-Grained Spatial Detail Without Increasing Bitrate

In real-world acoustic scenes, low-frequency ambience and high-frequency sound events arrive from different directions simultaneously. Preserving this spatial clarity requires high frequency resolution—something conventional low-bitrate parameter coders struggle to sustain.

Figure 2: Reconstruction of the 3-D directivity vector across ERB frequency bands at 750 bps comparing Ground Truth, Baseline DirAC, and Proposed MetaFOA.
Figure 2. Reconstruction of the 3-D directivity vector ([vt,b]y component) on a real-world STARSS23 recording at Rsp = 750 bit/s. As the number of frequency bands B increases from 3 to 36, MetaFOA progressively recovers sharper time-frequency spatial structure while reducing directivity distortion. PDF ↗
01

Natural Diffuseness Weighting

Each band’s 3-D directivity vector vt,b = √(1 − Dt,b) Ωt,b points in the direction of arrival while its length reflects directional energy. Diffuse background noise naturally shrinks toward the origin, so quantizer capacity is automatically focused on clear directional sources.

02

Predictable Bitrate, Flexible Quality

MetaFOA maintains a constant spatial metadata budget as the scene changes. The same quantizer scales from 0.75 to 3.0 kbps by selecting more RVQ stages, offering a straightforward trade-off between transmission rate and reconstruction quality.

03

36-Band Resolution at Zero Extra Bitrate

Because Cross-Band RVQ quantizes the full B × 3 directivity matrix as a unified pattern per frame, the spatial bitrate is completely decoupled from B. Scaling up to all 36 ERB auditory bands costs zero additional bits while faithfully preserving the ground-truth spatial image.

Reconstruction Quality Across Diverse Acoustic Environments

We evaluate FOA waveform and spatial reconstruction across three benchmarks—SpatialVCTK (anechoic spatial speech), STARSS23 (real-world multi-source sound events), and MEIR (measured reverberant rooms). The results cover spectral fidelity, spatial accuracy, and speech content preservation at several operating points. With spatial metadata starting at 0.75 kbps, MetaFOA supports fine frequency resolution and pairs flexibly with pretrained EnCodec or DAC monaural backbones.

Table 1. FOA reconstruction performance comparing Baseline DirAC [1] and Proposed MetaFOA. Lower values indicate better performance across all metrics. Best results per metric are shown in bold, and best results at Rsp = 0.75 kbps are underlined.
Method W-ch Codec Rsp
(kbps)
Stages
(S)
Rtot
(kbps)
SpatialVCTK (Speech) STARSS23 (Real Sound Events) MEIR (Reverberant Rooms)
STFT ↓ Mel ↓ Ang ↓ Diff ↓ WER ↓ STFT ↓ Mel ↓ Ang ↓ Diff ↓ STFT ↓ Mel ↓ Ang ↓ Diff ↓
Baseline (DirAC [1]) EnCodec 0.75 – 6.75 2.06 2.21 29.21° 0.00 0.15 2.04 1.67 47.42° 0.30 2.24 1.45 31.22° 0.24
EnCodec 1.50 – 7.50 1.18 0.94 3.56° 0.00 0.15 1.92 1.60 39.19° 0.32 1.97 1.31 18.24° 0.25
DAC 0.75 – 6.75 1.91 2.02 29.21° 0.00 0.14 1.66 1.22 47.48° 0.30 2.17 1.42 31.16° 0.24
DAC 1.50 – 7.50 0.91 0.60 3.56° 0.00 0.14 1.53 1.13 39.19° 0.32 1.87 1.28 18.15° 0.25
Proposed (MetaFOA) EnCodec 0.75 5 6.75 1.23 0.97 3.51° 0.02 0.15 1.65 1.44 37.98° 0.20 1.78 1.24 20.84° 0.22
EnCodec 1.50 10 7.50 1.22 0.95 2.72° 0.02 0.15 1.65 1.43 34.89° 0.20 1.78 1.24 19.61° 0.22
EnCodec 3.00 20 9.00 1.21 0.94 2.18° 0.02 0.15 1.64 1.43 33.06° 0.20 1.78 1.24 18.84° 0.23
DAC 0.75 5 6.75 0.99 0.67 3.51° 0.02 0.14 1.26 0.96 37.80° 0.20 1.69 1.21 20.76° 0.22
DAC 1.50 10 7.50 0.97 0.65 2.72° 0.02 0.14 1.26 0.96 34.61° 0.20 1.68 1.21 19.51° 0.22
DAC 3.00 20 9.00 0.96 0.63 2.17° 0.02 0.14 1.25 0.95 32.73° 0.20 1.68 1.20 18.75° 0.23