【正文】 d had not been very careful in delineating training and testing sets. As a result, it was very difficult to pare performance across systems, and a system39。Speech Recognition Victor Zue, Ron Cole, amp。s performance typically degraded when it was presented with previously unseen data. The recent availability of a large body of data in the public domain, coupled with the specification of evaluation standards, has resulted in uniform documentation of test results, thus contributing to greater reliability in monitoring progress (corpus development activities and evaluation methodologies are summarized in chapters 12 and 13 respectively). Finally, advances in puter technology have also indirectly influenced our progress. The availability of fast puters with inexpensive mass storage capabilities has enabled researchers to run many large scale experiments in a short amount of time. This means that the elapsed time between an idea and its implementation and evaluation is greatly reduced. In fact, speech recognition systems with reasonable performance can now run in real time using highend workstations without additional hardwarea feat unimaginable only a few years ago. One of the most popular, and potentially most useful tasks with low perplexity (PP=11) is the recognition of digits. For American English, speakerindependent recognition of digit strings spoken continuously and restricted to telephone bandwidth can achieve an error rate of % when the string length is known. One of the best known moderateperplexity tasks is the 1,000word socalled Resource 5 Management (RM) task, in which inquiries can be made concerning various naval vessels in the Pacific ocean. The best speakerindependent performance on the RM task is less than 4%, using a wordpair language model that constrains the possible words following a given word (PP=60). More recently, researchers have begun to address the issue of recognizing spontaneously generated speech. For example, in the Air Travel Information Service (ATIS) domain, word error rates of less than 3% has been reported for a vocabulary of nearly 2,000 words and a bigram language model with a perplexity of around 15. High perplexity tasks with a vocabulary of thousands of words are intended primarily for the dictation application. After working on isolatedword, speakerdependent systems for many years, the munity has since 1992 moved towards verylargevocabulary (20,000 words and more), highperplexity (PP≈200), speakerindependent, continuous speech recognition. The best system in 1994 achieved an error rate of % on read sentences drawn from North America business news. With the steady improvements in speech recognition performance, systems are now being deployed within telephone and cellular works in many countries. Within the next few years, speech recognition will be pervasive in telephone works around the world. There are tremendous forces driving the development of the technology。有些系統(tǒng)要求發(fā)言者登記 —— 即用 戶在使用系統(tǒng)前必須為系統(tǒng)提供演講樣本或發(fā)言底稿，而其他系統(tǒng)據(jù)說(shuō)是獨(dú)立揚(yáng)聲器，因 為沒(méi)有必要登記。 語(yǔ)音識(shí)別是一個(gè)困難的問(wèn)題，主要是因?yàn)榕c信號(hào)相關(guān)的變異有很多來(lái)源。數(shù)字化語(yǔ)音信號(hào)先轉(zhuǎn)換 成一系列有用 的測(cè)量值或有特定速率的特征，通常每次間隔 10 20毫秒（見(jiàn)第 章節(jié)，分別描述了模 擬信號(hào)和數(shù)字信號(hào)的處理）。 字級(jí)差異可以由發(fā)音網(wǎng)絡(luò)中可描述的字詞的候選發(fā)音來(lái)處理。 二 目前發(fā)展現(xiàn)狀 討論目前的發(fā)展?fàn)顩r ,需要聯(lián)系到具體應(yīng)用的環(huán)境 ,他影響到了任務(wù)的制約性。 首先， HMM 時(shí)代即將到來(lái)。 十年前，研究人員僅測(cè)試他們的 系統(tǒng)培訓(xùn)和利用當(dāng)?shù)厥占臄?shù)據(jù)，并沒(méi)有很仔細(xì)劃分培訓(xùn)和測(cè)試。對(duì)于美國(guó)英語(yǔ)，獨(dú) 立演講者的連續(xù)數(shù)字串識(shí)別和電話寬帶限制的語(yǔ)音可以達(dá)到 ％的誤碼率，前提是字符 串的長(zhǎng)度已知。 隨著語(yǔ)音識(shí)別性能的不斷改善，系統(tǒng)現(xiàn)正部署在電話和許多國(guó)家的蜂窩網(wǎng)絡(luò)。他們的表現(xiàn)可以得到進(jìn)一步加強(qiáng)，如果可以報(bào)考，如支配的具體領(lǐng)域限制的醫(yī) 療報(bào)告。目前，當(dāng)系統(tǒng)時(shí)常遭受重大 退化時(shí)，它便移動(dòng)到一個(gè)新的任務(wù)上。 超綱詞匯： 系統(tǒng)設(shè)計(jì)使用一套特定的單詞，但系統(tǒng)的用戶可能不知道哪些詞是屬于詞匯系統(tǒng)中 的。 如何把韻律信息整合到識(shí)別系統(tǒng)中來(lái)是一個(gè)尚未解決的關(guān)鍵性問(wèn)題。但據(jù)了解，對(duì)于文字和音素 知覺(jué)線索的性質(zhì)，其所需要整合的功能，反映了音節(jié)的動(dòng)態(tài)，這是動(dòng)態(tài)性的變動(dòng)整合。系統(tǒng)必須有一些方法來(lái)檢 測(cè)超綱的詞匯，否則最終將會(huì)從詞匯單詞映射到未知的單詞，導(dǎo)致發(fā)生錯(cuò)誤。 適應(yīng)： 如何能 適應(yīng)系統(tǒng)不斷變化的條件（新?lián)P聲器，麥克風(fēng)，任務(wù)等）和使用，通過(guò)使用改 進(jìn)？這種適應(yīng)可能發(fā)生在多層次的系統(tǒng)，模型子字，詞的發(fā)音，語(yǔ)言模型等。在語(yǔ)料庫(kù)的總機(jī)電話 交談字識(shí)別率是 50％左右。在未來(lái)幾年中，語(yǔ)音識(shí)別的電話網(wǎng)絡(luò)將在世界各地普 遍存在。最好的獨(dú)立執(zhí)行任務(wù)的語(yǔ)音設(shè)備執(zhí)行 RM 任務(wù)不超過(guò) 4％，用文字語(yǔ)言模型約束給定的單詞。公共 領(lǐng)域最近提供的數(shù)據(jù)按照評(píng)價(jià)標(biāo)準(zhǔn)的規(guī)范，致使試驗(yàn)結(jié)果相同，從而有助于提高監(jiān)測(cè)的可 靠性（語(yǔ)料庫(kù)發(fā)展活動(dòng)的主體和評(píng)價(jià)方法，分別在 12 和 13 章作了總結(jié)）。 第二，很大的努力已經(jīng)投入到語(yǔ)音系統(tǒng)大量詞匯識(shí)別的發(fā)展、訓(xùn)練和測(cè)試上。 例如，當(dāng)詞匯量小，整個(gè)單詞可以建模為一個(gè)單元。統(tǒng)計(jì)語(yǔ)言的模型 基于對(duì)字序列的發(fā)生頻率的估計(jì)，常常通過(guò)可能的詞序來(lái)引導(dǎo)搜索。 整個(gè)過(guò)程中，訓(xùn)練數(shù)據(jù)是用來(lái)確定模型 參數(shù)值的。 這些語(yǔ)音的變異性正好由音素的聲學(xué)差異做出了驗(yàn)證 。 當(dāng)詞匯量比較大或有較多象聲詞的 時(shí)候，識(shí)別起來(lái)一般比較困難。T, on the other hand, has installed a call routing s