Across the 400 records in this dataset, the leading assignee holds 19 records, with the top 5 assignees combined holding 78 records, or 19.5% of the total. That is meaningful concentration but not control — the top 10 together reach only 33.8%, leaving roughly two-thirds of filings spread across a long tail of smaller and single-filing entrants. Consumer electronics, cloud platforms, hearing-device makers and telecom licensing firms all appear among the more active filers, reflecting how broadly this technology has been absorbed into different product categories.

No — filing peaked at 38 records in 2022 and has declined since, with only 6 records in the most recent, still-partial year. Because publication typically lags actual filing by around 18 months, the final year or two of any trend understates true activity, so some of that drop will fill in. Even accounting for that lag, the multi-year decline from the 2022 peak suggests filing has cooled from its high point rather than continuing to accelerate.

This dataset is built on G10L21, H04R3 and G10L25, and the resulting records show G10L (speech and audio analysis/synthesis) present in 89.5% of all 400 records, making it the dominant class by a wide margin. H04R (loudspeakers and transducers) follows at 32.0% and H04M (telephonic communication) at 18.5%. AI-model computing under G06N appears in only 10.0% of records, indicating that while learned models are represented, the field is still primarily classified as traditional signal processing rather than machine-learning-first.

The clearest gaps sit adjacent to the dense G10L and H04R core rather than outside it entirely: wind-noise-specific sub-band modeling, low-complexity double-talk detection, and perceptual-evaluation-driven distortion tuning all show comparatively thin claim density relative to the overall field. Crossover into pictorial or video communication (H04N) also remains light, at just 1.8% of records, despite the growth of video conferencing hardware. Any of these branches warrants a dedicated freedom-to-operate check before drafting, since low density in a search corpus does not guarantee no blocking art exists.

WO2026049768A1, filed by Google and published in March 2026, claims a specific two-level sub-band architecture that separates wind-noise estimation from intelligibility restoration, using what the filing calls a specformer model at each stage. It does not block noise suppression broadly; it blocks close implementations of that two-stage sub-band structure and the specific non-linear modeling approach described. Teams working on wind-noise handling should review its claim scope carefully, but single-stage or full-band architectures, or different model families, sit outside its literal claims based on the abstract alone.