Fastest
Eureka on the web
When you want the answer in the next five minutes.
The agent works the prompt against patents and technical literature, citing every source.
Run your analysis now →https://eureka.patsnap.com/resources/blog/rd-blog/attention-based-encoder-decoder-speech-recognition-model-patent-landscape/ · Patsnap · data cut-off 2026-10-15 · downloaded from the live page
Patent Landscape · AI & Machine LearningAttention-Based Encoder-Decoder Speech Recognition Patents
65 patent families cover attention-based encoder-decoder speech recognition; SoftBank Group leads with 11 families, most filers hold two or fewer.
182
Published Records
65
Simple Patent Families
60%
In Force at Cut-off
US
Leading Jurisdiction
182 published records collapse into 65 simple (DOCDB) families; the difference is the same invention filed in more than one country. In force = 110 of 182 records still granted and maintained at the data cut-off; the rest are pending, withdrawn, refused or lapsed.
- 65 simple patent families and 182 published records sit behind this search, with no date limit applied to the search string itself.
- SoftBank Group Corp is named on 11 of the 65 families, the highest share of any single row at 16.9%.
- 30 of the 35 listed filers hold 2 or fewer families each, a long tail alongside the single largest row.
Published byPatsnap Research··6 min readSourced from Patsnap Eureka
Overview
What this dataset covers
This landscape covers attention-based encoder-decoder speech recognition, built from a search for records whose title, abstract or claims combine an encoder-decoder or sequence-to-sequence architecture with attention and speech recognition or ASR terminology. The search string carries no date limit, so the 65 simple patent families and 182 published records in scope reflect whatever matches that phrasing regardless of when it was filed. G10L, the speech and audio analysis and synthesis subclass, appears on 76.9% of the families, with G06N (AI model computing) and G06F (electric digital data processing) also present on a substantial share of the set.
65 families is a small pool by any measure. The figures here are a starting point for a more targeted search, not a map of the field: absence of a company, a classification or a country from this page means it did not match this particular search string, not that it is absent from the technology.
Families and records in scope
Ordered by volume only, not by importance or claim strength. Counted in simple (DOCDB) patent families, so one invention filed in several countries counts once. Names come from Patsnap's standardised assignee field as recorded on each document; subsidiaries and group companies are not merged. The icon before each name is read from the name itself (words such as University, Institute or Ltd); the data carries no organisation-type field, so treat it as a reading aid rather than a classification. Filings held by named individuals rather than organisations are removed where they can be identified (20 here); that identification is not exhaustive, so an individually held filing can still appear below. Either way those filings stay inside the totals above. This table shows the top 30 of the 35 filers returned for this search; the data endpoint returns at most 100 entities, so even that 35 is not the complete set of filers. A single family can name more than one filer, so these counts overlap and do not add up to the total.
Source: Patsnap Eureka. Assignee ranking and totals. Derived from one Patsnap search for Attention-Based Encoder-Decoder Speech Recognition Model Patent Landscape, covering all matching records regardless of filing date (the search string sets no date limit), retrieved 2026-10-15. Counts cover published records only and move as new filings publish.Run this in Eureka MCP
Data
Filing activity and technology split
Filing counts below are published records, not families, and are shown by earliest filing year; recent years are still publishing and are understated. Classification shares are read against all 65 simple patent families and can sum to more than 100% because a family can carry several IPC classes.
Filing trend by earliest filing year
Published records run from 2 in 2017 to a peak of 49 in 2021, with 14 recorded so far for 2026. 2022 onward is still filling in as later filings publish, so the most recent years understate eventual totals and no acceleration or slowdown should be read from them.
Counted in published records by earliest filing year, not de-duplicated into families. A filing is counted in the year it was filed, but it only becomes visible after it publishes: roughly 18 months for most routes, and up to 30 months where a PCT application enters national phase. 2022 onwards are therefore still filling in and are shaded; no growth figure on this page uses them. Because this series counts documents rather than inventions, a rise can mean more inventions or the same inventions filed in more countries; this dataset cannot separate the two.
Technology composition (IPC subclass)
G10L, the speech and audio analysis and synthesis subclass, is named on 50 of the 65 families (76.9%). G06N (AI model computing) and G06F (electric digital data processing) follow at 36.9% and 32.3% respectively; the chart lists 16 subclasses in total, with the largest few shown here.
Shares are the percentage of the 65 patent families in scope. One patent family can carry several IPC classes, so the shares add up to more than 100%.
Source: Patsnap Eureka. Filing trend and technology composition. Derived from one Patsnap search for Attention-Based Encoder-Decoder Speech Recognition Model Patent Landscape, covering all matching records regardless of filing date (the search string sets no date limit), retrieved 2026-10-15. Counts cover published records only and move as new filings publish.
Go deeper on Attention-Based Encoder-Decoder Speech Recognition Model Patent Landscape with Eureka
This page is one run against one query. Ask Eureka your own question about attention-based encoder-decoder speech recognition model patent landscape and every answer comes back with the patent numbers behind it.
Try EurekaExample Record
A record from the result set
Example record
Sequence-to-sequence speech recognition with latency threshold
MICROSOFT TECHNOLOGY LICENSING, LLC
A computing system including one or more processors configured to receive an audio input. The processors generate a text transcription at a sequence-to-sequence speech recognition model, which assigns external-model text tokens to frames in the audio input, each with an external-model alignment. Based on the audio input, the system generates hidden states, and based on those hidden states, a set of output text tokens, each with a corresponding alignment.This is the first record returned under the database's default ordering, not a ranking by citation, breadth or date. It is shown to illustrate what one family in this dataset actually claims.
Selected as the first record the search returns under the database default order, then shown in full. It is not ranked by citations, claim breadth or commercial importance, and it is not a judgement about which filing in this space matters most.
Source: Patsnap Eureka. The first record the search returned, shown in full. Derived from one Patsnap search for Attention-Based Encoder-Decoder Speech Recognition Model Patent Landscape, covering all matching records regardless of filing date (the search string sets no date limit), retrieved 2026-10-15. Counts cover published records only and move as new filings publish.Run this in Eureka MCP
Run it yourself
Put your own technology through the same analysis
Where to run it
For builders
MCP server & REST API
When it has to run inside your own pipeline.
Patent search, landscape analysis and assignee resolution as MCP tools. Drop them into any agent framework, or call REST directly.
Browse MCP servers →Patterns
What the numbers show
Three figures from this dataset are worth sitting with before treating the page as a map of the field.
Scope
65 families / 182 records
simple patent families / published records
Two different counts, two different units
The dataset in scope totals 65 simple patent families and 182 published records. The search string carries no date filter, so this reflects whichever matching documents exist, filed whenever they were filed; the chart's recent years are still publishing.
Counted at retrieval on 2026-10-15.
Filer
11 of 65 families (16.9%)
largest single-row share
SoftBank Group Corp's single-row share
SoftBank Group Corp is named on 11 of the 65 families in scope, the highest figure of the 35 listed filers. Filer rows are not rolled up to corporate group and one family can name several filers, so this figure belongs to this one row only and cannot be added to any other row.
35 organisations listed; the table shows the top 30.
Distribution
30 of 35 filers ≤ 2 families
long tail of listed filers
Most listed filers hold very few families
30 of the 35 listed filers hold 2 or fewer simple patent families each. That leaves a small number of rows, including the 11-family leader, accounting for more filings than the rest of the list individually, without implying any combined share across rows.
Individuals are excluded from this list but remain in the 65-family total.
Legal status
110 of 182 records (60%)
granted and in force at retrieval
Status is a snapshot, not a success rate
110 of the 182 published records were granted and still in force at the 2026-10-15 retrieval date. The remaining records cover several other statuses that are not broken out individually here, and the figure describes legal status at one moment, not quality or enforceability.
Status changes continuously; this is a point-in-time read.
Eureka AI Agent
Looking for what nobody has claimed yet?
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to attention-based encoder-decoder speech recognition model patent landscape, with the prior art for and against each one.
Source: Patsnap Eureka. Derived observations on the same dataset. Derived from one Patsnap search for Attention-Based Encoder-Decoder Speech Recognition Model Patent Landscape, covering all matching records regardless of filing date (the search string sets no date limit), retrieved 2026-10-15. Counts cover published records only and move as new filings publish.Run this in Eureka MCP
Next steps
Where to take this from here
This page is a starting point built from one search string. The following moves narrow or widen the picture depending on what a reader needs next.
Read the granted claims, not the abstract
A family count or an abstract cannot establish what any patent actually blocks. Pull the granted claim set for any record that matters before drawing a freedom-to-operate conclusion.
Open the record in EurekaRe-run the search with adjacent terms
A single search string over- and under-captures by construction. Related phrasing around transformer-based ASR, CTC decoding or streaming attention may surface families this string did not catch.
Build a new search in EurekaCheck filer rows individually, not combined
Because filer rows overlap and are not rolled up to corporate group, any group-level reading of concentration needs to be built row by row rather than by summing the list.
Explore filers in EurekaSource: Patsnap Eureka. Suggested next steps. No new data: every figure named here is one of those above. Derived from one Patsnap search for Attention-Based Encoder-Decoder Speech Recognition Model Patent Landscape, covering all matching records regardless of filing date (the search string sets no date limit), retrieved 2026-10-15. Counts cover published records only and move as new filings publish.Run this in Eureka MCP
FAQ
Common questions
This particular search returns 65 simple patent families and 182 published records, with no date limit applied to the search string. A simple family counts one invention once no matter how many countries it was filed in, while published records count every publication, so the same invention filed in several countries produces several records. Because the search string carries no date limit, this is whatever currently matches the phrasing, not a fixed historical span, and a different phrasing would likely return a different count.
SoftBank Group Corp is named on 11 of the 65 simple patent families in scope, the highest figure among the 35 listed organisations. 30 of those 35 listed filers hold 2 or fewer families each, so the list has one row with a noticeably higher count and a long run of rows with very few. Filer rows are not rolled up to corporate group, and because a single family can name more than one filer, these rows cannot be added together into a combined share.
The record describes a computing system that receives audio input and generates a text transcription using a sequence-to-sequence speech recognition model, which assigns text tokens to frames of the audio with corresponding alignments, generating hidden states and output tokens from them. That description comes from the abstract alone, which is not a reliable guide to what the granted claims require or how far they reach. Anyone assessing freedom to operate against this record needs to read the granted claim set directly rather than inferring scope from the abstract or from any family count.
This page does not show a growth rate for this topic; the dataset is too thin and the recent years are still filling in to support one. The filing trend chart counts published records by earliest filing year, running from 2 records in 2017 up to a recorded peak of 49 in 2021, with 2022 onward still filling in as later filings publish. Reading a slowdown or acceleration into the most recent years would mistake publication lag for an actual change in filing behaviour.
G10L, the speech and audio analysis and synthesis subclass, is named on 50 of the 65 simple patent families (76.9%), the highest share of any subclass in this dataset. G06N, covering AI model computing, is named on 36.9% of families, and G06F, electric digital data processing, on 32.3%; a single family can carry more than one of these classes, so the shares add to more than 100%. The full classification picture spans 16 subclasses in total, of which these three carry the largest shares.
110 of the 182 published records in scope were granted and still in force at the 2026-10-15 retrieval date, which is 60% of the records. The remaining 40% cover several other statuses, including pending, withdrawn, refused and lapsed applications, which are not broken out individually in this dataset. This figure is a snapshot at one date rather than a measure of quality or a success rate, and legal status changes continuously as applications proceed.
Answers are grounded in the same dataset. Derived from one Patsnap search for Attention-Based Encoder-Decoder Speech Recognition Model Patent Landscape, covering all matching records regardless of filing date (the search string sets no date limit), retrieved 2026-10-15. Counts cover published records only and move as new filings publish.Run this in Eureka MCP
Research Attention-Based Encoder-Decoder Speech Recognition Model Patent Landscape in depth with Eureka
Go past this page: query the whole attention-based encoder-decoder speech recognition model patent landscape corpus yourself, in your own scope.
Every answer comes back with patent numbers you can open.
How this page was built, and what the numbers do and do not mean
Search. Every figure on this page comes from one Patsnap search, run once and not edited afterwards. This is the exact search string:
TAC:(("encoder-decoder" OR "sequence-to-sequence") AND attention AND ("speech recognition" OR ASR))Reading the field codes: TAC searches the title, abstract and claims of each record, not the full description. The search string sets no date limit, so the scope is every matching record in the database regardless of when it was filed. It is not restricted to recent years. How this string was built matters for how far the counts can be pushed. It was generated in batch from a topic list rather than hand-crafted for this page, and the result set is not manually screened. A keyword search over title, abstract and claims both over-captures (unrelated filings that happen to use the same words) and under-captures (the same invention written in different words). This page cannot measure either effect, so read every figure here as what this particular string returned, not as what exists in the field. One subclass (G10L) carries 76.9% of the patent families, so this search returned a classification-tight set.
What is being counted. Two different units appear on this page and they are not interchangeable. The filer list and the technology split are counted in simple (DOCDB) patent families: one invention counts once however many countries it was filed in. The total, the jurisdiction split and the filing-year chart are counted in published records: the same invention counts once per publication. For this search that is 182 records and 65 families. Record counts reward filing in many jurisdictions; family counts do not. Shares are only ever taken against a denominator in the same unit, and each one names it.
Who the filers are. Names come from Patsnap's standardised assignee field, as recorded on the documents themselves. They are not rolled up to corporate group: a parent and its subsidiaries appear as separate entries, and a portfolio acquired from another company stays under the name on the document unless the assignment is reflected there. This list is meant to carry organisations, not people. Filings held by a named individual are identified by matching the assignee against the inventors on the same search, and, for names written in Chinese, Japanese or Korean characters, by the name carrying no organisational suffix; those are removed from the list (20 of them here). Neither test is exhaustive: the inventor list it matches against is itself capped, and a personal name written in Latin script with no organisational wording in it cannot be told apart from a short company name by rule. So an individually held filing can still appear in the list. Those filings stay inside every total and share on this page either way, so the listed entries can add up to less than the total. A record shown in full further down still names its assignee exactly as the patent office records it, including where that is a person: that name is how the record is identified and checked. What is left out of the ranking is people as ranked entries, not people as cited sources. The table shows the top 30 of the 35 filers returned, ordered by count alone; the data endpoint returns at most 100 entities, so even that 35 is not every filer. One patent family can name several filers and each is counted separately, so these rows are not slices of a divided whole: adding them together over-counts, and this page deliberately publishes no combined "top five" share for that reason. See every filer, individuals included
Dates. The year chart counts by earliest filing year, not publication year. A filing only becomes visible once it publishes: roughly 18 months for most routes, and up to 30 months where a PCT application enters national phase. Measured across these pages, the year four back from today still holds only about 89% of its eventual count. 2022 onwards are therefore shaded on the chart as still filling in. The data behind this page was retrieved on 2026-10-15.
Legal status. "In force" counts documents whose legal status at retrieval was granted and maintained (110 of 182). Everything else is pending, withdrawn, refused, lapsed or not determinable, and those states are not broken out here because the underlying codes do not separate them cleanly. Legal status changes continuously and is reported per document, not per family. Get the status breakdown in Eureka
Classification. The technology split uses IPC subclasses (four characters, for example H01M). A single record usually carries several, so the shares add up to more than 100% and a record can appear in more than one bar. Subclass is a coarse grouping; it describes the field a record was classified into, not what the claims cover. Break the classes down to subgroup level
Who wrote the words. The tables, charts, counts, shares and patent numbers on this page are produced directly from the search result by a program; none of them are written or estimated by a language model. The surrounding prose, headings, FAQ answers and the answers in the question panel are drafted by a language model that is given those same figures and is not permitted to introduce others. Prose can still mis-describe a correct figure, so where the two disagree, the tables and charts are the record.
Disclaimer. This page is generated from Patsnap Eureka data drawn from a limited snapshot of global patent and scientific-literature records, and is provided for general information and reference only.
Patent data carries limitations that apply to every figure above. Recent filings are under-counted, because an application is not public until roughly 18 months after filing and database loading adds further delay. The search string defines and constrains everything shown: records that use different terminology, or that sit outside the fields searched, are absent and cannot be counted. Applicant names, classification codes, legal-status codes and citation records are taken as recorded and may contain errors, duplicates, omissions or stale entries. Ordering by volume is not a measure of quality, scope, enforceability or commercial value, and a larger portfolio is not a stronger one. As a result, the analysis may be incomplete or inaccurate and may not reflect the full technology landscape.
Nothing on this page constitutes an exhaustive prior-art, novelty, freedom-to-operate, validity or infringement search, nor does it constitute legal, financial, investment, or professional advice, and it should not be relied upon as such. No attorney-client relationship is created by it. Any patent, commercial, or strategic decision should be verified independently against the primary records and reviewed with qualified patent, legal, and domain professionals. Patsnap makes no warranties, express or implied, as to the accuracy, completeness, or fitness for any particular purpose of the information presented.
The question panel. The answers behind the suggested questions on this page were written when the page was generated, not in response to your click. The progress steps that play before an answer appears are an animation of how that analysis was done; nothing is being searched or computed at that moment. Typing your own question instead opens Patsnap Eureka, which does run it live.
Machine translation. Assignee and organisation names originally recorded in Chinese, Japanese or Korean have been rendered into English by an AI translation step so that the tables stay readable. These renderings are best-effort and may not match a company's registered English name; the original name is what the underlying patent record carries, and it is what any Eureka query launched from this page uses.