Bartłomiej Mróz

This paper presents a neural network for binaural rendering of first-order Ambisonics (FOA) signals, enabling immersive spatial audio over headphones without individualized HRTF measurements at inference time. The model operates in the STFT domain using Complex Ratio Masks (CRM): a shared mask pair (left/right ear) is applied via complex-valued multiplication to all four FOA channels with directional weighting, preserving both magnitude and phase. The input representation extends standard spectral features with acoustic intensity vector channels encoding sound arrival direction at each time-frequency bin. Training uses a multi-objective loss combining SI-SDR, multi-resolution spectral reconstruction, and interaural level/phase difference terms. The four-level UNet backbone with channel-spatial attention totals approximately four million parameters. Evaluation against eight baseline configurations spanning magnitude-masking variants (28M parameters), a waveform-domain Wave-U-Net (143 M), and the A2B time-domain renderer shows that the CRM variant achieves the strongest spectral and spatial fidelity in the cross-domain setting with a seven-fold parameter reduction over the magnitude-masking baseline while maintaining real-time CPU operation.

The Zylia ZM-1 (approximately 103 mm spherical array, 19 MEMS capsules, 3rd-order) and Harpex Spcmic (170 mm planar array, 84 MEMS capsules, 5th-order capable) represent two distinct geometrical approaches to higher-order Ambisonics capture. Despite widespread adoption, systematic comparison of their performance under real recording conditions is absent from published literature. We present a controlled case study based on simultaneous recordings of solo piano repertoire in a 370-seat concert hall (Aula Politechniki Gdańskiej, RT60 = 1.97 s), with both arrays mounted 17 cm apart on a single stereo bar. After peak-normalization to -0.5 dBTP, four metrics were compared: W-channel spectral response, integrated loudness (ITU-R BS.1770-5), per-order spatial energy distribution, and first-order directional ratios. The headline finding concerns spatial energy: at nominally identical 3rd-order, the ZM-1 attenuates 27.4 dB from 0th to 3rd order, while the Spcmic attenuates only 8.4 dB – a 19 dB gap that is not driven by aliasing-band content and persists across repertoire. The results indicate that nominal Ambisonics order alone does not predict per-order energy distribution in real recordings; geometry and encoding contribute substantially even when the format label is identical. Recordings are openly available under CC BY-NC-SA 4.0.

Spatial audio recording using higher-order Ambisonics offers rich directional information for medical speech capture, yet challenging hospital acoustic environments motivate preprocessing with neural denoising algorithms. This study investigates whether U-Net-based denoising of third-order ambisonic recordings improves automatic speech recognition (ASR) quality for medical applications. We developed the Medical Immersive Audio Corpus (MIAC), comprising 1,759 utterances (6.43 hours) of Polish medical speech recorded with a Zylia ZM-1 microphone in uncontrolled hospital environments, capturing 16-channel third-order Ambisonics across multiple specializations including thyroid ultrasonography, surgical procedures, and general diagnostics. We applied a U-Net architecture with dual attention mechanisms trained using the Noise2Noise paradigm to denoise the corpus, then evaluated transcription quality using ten Whisper ASR models ranging from 39 million to 1.55 billion parameters, including domain-adapted medical variants. Surprisingly, we discovered a noise reduction paradox where denoising degraded transcription quality for seven of ten models, with statistically significant increases in Word Error Rate (WER) and Character Error Rate (CER) for general-purpose base, small, and medium models. Only the domain-adapted whisper-medium-68000-abbr model showed statistically significant improvement (p = 0.0008), while large-scale models (large-v2, large-v3) exhibited robustness with negligible changes. Effect sizes remained small (Cohen’s d < 0.2) across all models. These counterintuitive findings suggest modern ASR systems implicitly utilize background noise characteristics as informative features, and that preprocessing pipelines should be reconsidered for domain-specific applications. Our results provide practical guidance for medical speech processing system design.

This paper presents a novel audio decorrelation method that integrates velvet noise with parametric modeling of vowel filters derived from recorded speech. By capturing vocal timbre, the technique extends decorrelation and artificial reverberation tools, offering new creative possibilities. Velvet noise is filtered to model vowel resonances, enabling users to imprint speech or singing characteristics onto multichannel effects. Inspired by choir acoustics, the system synthesizes distinct vowels per channel, producing speech-like textures that are convolved with input audio. The approach emphasizes immersive, voice inspired sound design, with discussion of implementation challenges, creative applications, listening tests, and directions for future research.

Immersive audio formats compete for adoption, yet objective comparisons across them remain limited, and the behaviour of the objective models under cross-format conditions is largely uncharacterised. This study applies two binaural models, BAM-Q (perceptual quality) and BINAQUAL (localization similarity), to binaural renderings of 11 format variants derived from 3 professional recordings differing in genre and Ambisonic provenance, under three reference anchors and with a controlled test of stimulus alignment. Two properties of the evaluation influence the outcome more strongly than many format differences, and affect the two models differently. First, immersive renderers impose transport delays up to 510 ms that are not quality degradations; because BINAQUAL compares phase representations, leaving them uncompensated makes its score track delay rather than fidelity (ρ = -0.86), and correcting them changes localization similarity by a median of 149.3%. BAM-Q, operating on modulation-domain representations, is unaffected (median 0.0%). Second, under BINAQUAL the apparent advantage of one format family over another follows the anchor’s family, reversing in all 3 items when a channel-based anchor replaces a scene-based one; the BAM-Q composite never reverses. With aligned stimuli the models are moderately correlated (r = 0.51), not independent. Format orderings are reproducible across content (Kendall’s W = 0.94 for localization similarity, permutation p < 0.001). The results yield prescriptions for cross-format evaluation (compensate and report latency, treat the anchor as an experimental factor, refer cue measures to discrimination thresholds) and identify subjective validation as the next step.

A containerized, low-latency higher-order Ambisonics (HOA) and 360-degree video streaming pipeline exhibited a persistent ~2-s audio–video desynchronization, isolated to Chromium-based browsers. The audio path was the natural suspect – wrongly, for weeks. The actual fault was a leading empty edit (elst, media_time = -1) in the fragmented MP4 stream, written whenever contributed audio and video begin at slightly different times, a common live-ingest occurrence. Each of the three major engines disagrees with itself: Chromium and Safari apply the offset in progressive playback but drop it under Media Source Extensions (MSE), while Firefox reverses. I present a minimal, reproducible test case, filed as a cross-engine, specification-level question against Chromium, WebKit, Firefox, and the W3C standard. WebKit has landed a fix; Chromium has stated intent to follow; Firefox has triaged the report. I also report a source-side fix and a client-side workaround. The lesson: a difference measurement proves disagreement, never which side is wrong.

The ventriloquist effect, understood as perceptual displacement of sound toward a spatially discrepant visual stimulus, has been extensively studied in traditional laboratory settings, but its manifestation in immersive virtual reality with naturalistic movement remains poorly understood. We investigated audiovisual spatial binding in a six-degrees-of-freedom (6DoF) VR environment using fifth-order Ambisonics spatial audio. Thirty-one participants localized sounds presented with visual cues displaced by 15-70 cm. Results revealed a robust ventriloquist effect: participants’ localization responses were shifted approximately 35% toward the visual cue on average, with substantial individual differences (range: 11-73% visual capture). Sound type did not modulate the effect, but exploratory analyses revealed that participants who moved more during trials achieved better localization accuracy, suggesting that head movement in 6DoF VR serves as a beneficial exploration strategy. We introduce ASCOE (Auditory Spatial & Cognitive Observation Environment), an open-source VR research platform for studying audiovisual perception, and provide complete analysis scripts for reproducibility. These findings extend classic ventriloquist research to ecologically valid immersive contexts and have implications for spatial audio design in VR applications.

Research in spatial audio quality assessment, spatial machine learning, and immersive production is constrained by the scarcity of diverse, well-documented Higher-Order Ambisonics (HOA) recordings: existing public datasets are dominated by first-order, synthetic, or speech-only material. This paper presents an openly available seven-year HOA recording corpus comprising 23 sessions captured between April 2019 and January 2026 across concert halls, churches, studios, and outdoor environments, with musical content ranging from solo piano to 60+ performer ensembles. The corpus was captured primarily with a Zylia ZM-1 (19 MEMS capsules, 3rd order), with comparative material from a Harpex Spcmic (84 capsules, 5th order) and a Saramonic SR-VRMIC (4 electret capsules, 1st order); all recordings are provided in B-format using the AmbiX (ACN/SN3D) convention. To demonstrate the corpus’s research utility, we report a co-located comparison of the three arrays under matched acoustic conditions, which reveals that nominal Ambisonics order alone does not characterize spatial resolution: the ZM-1 exhibits 27.4 dB of energy roll-off from 0th to 3rd order, against only 8.4 dB for the Spcmic at the same order – a difference consistent across two independently performed works and large relative to its bootstrap confidence interval. We document the full recording and signal-processing methodology, measured room-acoustic parameters for the primary venue (RT60 = 1.97 s), and integrated-loudness quality metrics for all 74 rendered files. The corpus, metadata, and a reproducible analysis pipeline are released under a CC BY-NC-SA 4.0 license to support further research in spatial audio quality assessment, compression-algorithm evaluation, and immersive production.

In this article spatial audio recording techniques are compared: scene-based audio and object-based audio. The study involved mixing recordings from a higher-order ambisonic microphone and support microphones, ambisonically encoded on a virtual sphere. The recordings were combined in different spatial resolution variations by manipulating the ambisonic order. A MUSHRA-like test was conducted, taking into consideration the room divergence effect. The experiment used binaural rendering with headtracking. The results were analyzed using linear mixed models, providing insights into spatial audio recording techniques.