研析語音辨識技術應用於食品販售相關網路媒體之探索   全文下載

Exploration of the Application of Speech Recognition Technology in Food Sales-related Online Media

隨著網路購物風潮的興起，人們越來越依賴網路平台購買食品，特別是在網路直播和影片分享平台上，許多影片內容會介紹餐廳美食心得，或者是與業者合作，推廣其產品並提供購買方式，這些影片往往對消費者產生重大影響，因此，網路食品安全管理應納入整體風險管控的範疇中。本次透過語音辨識模型測試三種不同網路食品販售類型影片，發現目前技術仍存在一些缺點。如在有吵雜的背景聲情況下，辨識結果受到明顯影響，僅能辨識出零星字詞而無法提供完整語句。其次，對於音訊檔中包含專有名詞的情況，容易發生辨識錯誤，尤其是當處理特殊領域的專有名詞時，另人聲講話時若字詞不清晰且有連音現象，可能導致對單字的辨識錯誤，或在多人同時講話的場景中，辨識系統傾向於取人聲相對較大的部分進行辨識，而忽略人聲較小的部分，可能造成對話內容的不完整或錯誤辨識。以上這些缺點凸顯了目前在處理複雜音訊環境和特殊語音情境時，仍有精進的空間，在未來應用中，仍需要克服上述缺點，根據具體需求調整相應參數及辨識檔案質量，以提升辨識準確性。

With the rise of online shopping trends, people are increasingly relying on internet platforms to purchase food, particularly through livestreams and video-sharing platforms where content often includes restaurant reviews or collaborations with vendors promoting their products with purchasing options. Such videos wield significant influence over consumers, thus necessitating the inclusion of online food safety management into the overall risk mitigation strategies. A study testing three different types of online food sales videos using speech recognition models revealed several shortcomings in current technology. In scenarios with background noise, recognition results were noticeably affected, often yielding sporadic words instead of complete sentences. Additionally, errors in recognizing proprietary terms were common, particularly when dealing with specialized fields. Furthermore, unclear speech or instances of connected speech led to recognition errors, and in multi-speaker settings, the system tended to prioritize louder voices, potentially resulting in incomplete or incorrectly recognized dialogue. These limitations underscore the need for improvement in handling both complex audio environments and specialized speech contexts. Future applications shall overcome these challenges by adjusting relevant parameters and enhancing recognition file quality to improve accuracy.