# iFLYTEK speech recognition: matching meeting terminology to the audio

How EP4156176B1 combines two attention stages, contextual terms and dynamic scoring, with a clear distinction from service input limits.

Source: https://www.iplexlaw.co.kr/en/blog/iflytek-contextual-hotword-speech-recognition

HOME / NEWS & INSIGHTS NEWS & INSIGHTS iFLYTEK speech recognition: matching meeting terminology to the audio How EP4156176B1 combines two attention stages, contextual terms and dynamic scoring, with a clear distinction from service input limits. Artificial Intelligence 2026.11.14 published IPLEX 6 min read Contextual terminology supplies candidates without replacing audio evidence. A meeting transcript may capture ordinary sentences while confusing a crucial product code. The iFLYTEK and University of Science and Technology of China patent uses a contextual vocabulary during decoding: terms help locate useful audio, and that audio helps distinguish the terms. 1. Why a short code matters Imagine a prototype review using AX-17 and AX-70. A wrong suffix can send the team to the wrong test report even when the surrounding sentence is accurate. An acoustic model has less familiarity with company-specific names than everyday expressions. Shared sounds alone cannot reliably distinguish the two codes. The joint applicants’ European patent EP4156176B1 was published as granted on 10 September 2025. It describes consulting a prepared vocabulary while converting speech to text. 2. Encode speech and context separately The system first encodes the audio and the candidate hotwords separately. Here, hotword means a contextually relevant term, not a command that necessarily triggers an assistant. A meeting-specific list supplies possible names or abbreviations; it does not establish that any of them was actually spoken. Audio evidence remains necessary. The list participates in decoding rather than merely replacing strings in a finished transcript. This distinction matters when describing the technical processing sequence. 3. Two attention stages exchange evidence The first attention model uses the hotword encoding and the decoder state, which represents preceding decoded information, to select the relevant audio feature. The second attention model uses that audio feature to weight the candidate hotwords. The two resulting features are combined for decoding and classification. The classifier considers ordinary characters, dynamically added term candidates and the non-hotword case. Adding a word is therefore not an instruction to force it into the transcript. 4. Distinguishing AX-17 from AX-70 In the hypothetical meeting, the shared AX sound may bring both codes into consideration. The later sound and the current decoding context help distinguish their endings. The diagram illustrates how candidate terminology and audio can constrain one another; it does not report measured recognition accuracy or establish support for these codes in a commercial service. Testing must also include ordinary speech and expressions absent from the list. Excessive contextual bias can insert a listed term where the speaker said something else. 5. Patent mechanisms and actual products The referenced XFYUN real-time transcription documentation describes customized hotwords, but its standard input accepts UTF-8 Chinese characters and excludes digits and Latin letters. AX-17 is an explanatory example of the patent mechanism, not a supported entry for that service. Hypothetical AX-17 and AX-70 example illustrating the two-way comparison. iFLYTEK’s AINOTE 2 product page separately describes transcription, translation and meeting summaries. These functions do not establish that the product implements the particular patent claims. A mistaken code can propagate from transcription into translation or a summary. Evaluating the final prose alone can conceal the original identification error. 6. What to record and compare Scope and version the vocabulary for the meeting. Compare critical-code recovery as well as false insertions and ordinary-word recognition, rather than relying only on an aggregate error rate. For technical disclosure, record both encoders, the decoder-state inputs, the order of the attention stages and the scoring of ordinary and contextual candidates. Keep the proposed mechanism, the service’s documented input restrictions and any measured results distinct. A useful comparison tests how the same audio behaves with and without the contextual processing. The original figure labels and forms are preserved; the text explains the source material. The meeting and product codes are hypothetical; product descriptions do not establish implementation of specific patent claims. EPO patent EP4156176B1 XFYUN real-time transcription documentation iFLYTEK AINOTE 2 product information Socra AI’s adaptive-learning patent: different wrong answers, different next questions NVIDIA early pruning: when does a smaller network become stable enough? Explore IPLEX’s AI projects and consultation materials for speech recognition and contextual decoding. Explore our AI and patent practice Original Tistory article Read the Korean source This article reflects the information available when it was published. Contact us to discuss your circumstances. Discuss this topic ↗ All articles ON THIS PAGE 1. Why a short code matters 2. Encode speech and context separately 3. Two attention stages exchange evidence 4. Distinguishing AX-17 from AX-70 5. Patent mechanisms and actual products 6. What to record and compare TALK TO IPLEX Discuss your IP questions We consider your technology and business needs together. ↗ Contact us Newer When patent customs detention and warning letters do not establish liability ↗ Older Garment designs shown on mannequins: reproducibility and overall impression ↗ Related insights Artificial Intelligence 2026.11.19 Matryoshka embeddings for image retrieval: shortlist with a prefix, rerank with the full vector A practical reading of Matryoshka Representation Learning, its nested training objective and the ImageNet-1K retrieval results reported at NeurIPS 2022. ↗ Read article Artificial Intelligence 2026.11.18 ByteDance’s captioning patent: transcribing separate tracks and resolving overlapping display times How US11837234B2 processes selected audio tracks and adjusts caption timing, with a practical distinction between patent disclosure and CapCut’s public guidance. ↗ Read article Artificial Intelligence 2026.11.17 Apple recognized text: connect a phone number or web address to the right action US12001642B2 links recognized text, image coordinates, text types and user input to contextual menus and actions. ↗ Read article

- https://www.iplexlaw.co.kr/en
- https://www.iplexlaw.co.kr/en/blog/category/ai
- https://data.epo.org/publication-server/rest/v1.0/publication-dates/20250910/patents/EP4156176NWB1/document.pdf
- https://www.xfyun.cn/doc/asr/rtasr/rtasr-description.html
- https://www.iflytek.com/en/products/air/ainote-2.html
- https://www.iplexlaw.co.kr/en/blog/socra-ai-personalized-learning-distractor-tags
- https://www.iplexlaw.co.kr/en/blog/nvidia-early-structural-pruning-stability
- https://www.iplexlaw.co.kr/en/ai-patents
- https://iplexlaw.tistory.com/26
- https://www.iplexlaw.co.kr/ko/blog/iflytek-contextual-hotword-speech-recognition
- https://www.iplexlaw.co.kr/en/contact
- https://www.iplexlaw.co.kr/en/blog
- https://www.iplexlaw.co.kr/en/blog/iflytek-contextual-hotword-speech-recognition#article-26-en-0
- https://www.iplexlaw.co.kr/en/blog/iflytek-contextual-hotword-speech-recognition#article-26-en-1
- https://www.iplexlaw.co.kr/en/blog/iflytek-contextual-hotword-speech-recognition#article-26-en-2
- https://www.iplexlaw.co.kr/en/blog/iflytek-contextual-hotword-speech-recognition#article-26-en-3
- https://www.iplexlaw.co.kr/en/blog/iflytek-contextual-hotword-speech-recognition#article-26-en-4
- https://www.iplexlaw.co.kr/en/blog/iflytek-contextual-hotword-speech-recognition#article-26-en-5
- https://www.iplexlaw.co.kr/en/contact
- https://www.iplexlaw.co.kr/en/blog/brunch-case-1961
- https://www.iplexlaw.co.kr/en/blog/brunch-case-1960
- https://www.iplexlaw.co.kr/en/blog/matryoshka-prefix-vector-image-retrieval
- https://www.iplexlaw.co.kr/en/blog/matryoshka-prefix-vector-image-retrieval
- https://www.iplexlaw.co.kr/en/blog/bytedance-track-based-caption-timing
- https://www.iplexlaw.co.kr/en/blog/bytedance-track-based-caption-timing
- https://www.iplexlaw.co.kr/en/blog/apple-recognized-text-contextual-actions
- https://www.iplexlaw.co.kr/en/blog/apple-recognized-text-contextual-actions
