Across the 107,918 records in this dataset, filing is concentrated among a small set of large technology licensing entities, device makers and cloud platforms, with the leading assignee holding 4,216 records. The top five combined account for 16,348 records, or 15.1% of all records in scope, and the top ten reach 24.1%. That leaves a long tail of smaller filers holding the majority of the field, so no single company controls the claim space outright.

Yes, through the last complete filing year the trend was positive: annual filings rose from 1,330 in 2021 to 1,400 in 2024, a 5% increase, with 2023 marking the highest single year so far at 1,447 filings. The 2025 and 2026 figures in the dataset look lower, but that reflects publication lag — patents typically publish around 18 months after filing — rather than an actual slowdown in activity.

G10L, covering speech and audio analysis and synthesis, is the largest single class, appearing in 12.5% of all 107,918 records. G06F (electric digital data processing) follows at 5.4% and G06N (AI-model computing) at 3.9%, with image recognition, business-data processing and telephonic communication classes each below 1%. Because a single patent can carry several IPC classes, these percentages are each measured against the full record total and are not meant to sum to 100%.

This filing from Electronics and Telecommunications Research Institute claims a speech recognition apparatus that estimates reliability of a time-frequency segment in an input voice signal, then reflects that reliability into a normalized cepstrum feature vector and average vector used during HMM-based decoding, using a discrete cosine transformation to produce a transformed cepstrum vector. It is specific to that reliability-weighted cepstrum transformation approach rather than to speech recognition generally, so systems using different feature representations, different reliability weighting, or decoding methods other than HMM would not fall inside its scope on the same terms. Anyone building a cepstrum-based front end with confidence weighting should review its exact claim language before assuming freedom to operate.

Filing density is heavily weighted toward core acoustic classes like G10L, while adjacent branches such as cross-lingual feature vector transfer, low-resource speech signal augmentation, on-device inference compression, and multimodal speech-vision alignment show comparatively thinner coverage. That thinner coverage does not mean the technology is undeveloped, only that fewer claims currently occupy that space, which is where a well-drafted first claim is more likely to find open ground. Confirming this requires checking the specific claim scope of nearby filings, not just the class-level counts.