Historically, basic speech recognition research has focused almost exclusively on optimizing large vocabulary speaker-independent recognition of continuous dictation. A major impetus for this research has come from US government sponsored competitions held annually by the Defense Advanced Research Projects Agency (DARPA). The main emphasis of these competitions has been on improving the "raw" recognition accuracy--calculated in terms of average omissions, insertions, and substitutions--of large-vocabulary continuous speech recognizers (LVCSRs) in the task of recognizing read sentence material from a number of standard sources (e.g., The Wall Street Journal or The New York Times). The best laboratory systems that participated in the WSJ large-vocabulary continuous dictation task have achieved word error rates as low as 5%, that is, on average, one recognition error in every twenty words (Pallet, 1994).
In recent years, an increasing number of speech laboratories have begun deploying speech technology in CALL applications. Results include voice-interactive prototype systems for teaching pronunciation, reading, and limited conversational skills in semi-constrained contexts. Our review of these applications is far from exhaustive. It covers a select number of mostly experimental systems that explore paths we found promising and worth pursuing. We will discuss the range of voice-interactions these systems offer for practicing certain language skills, explain their technical implementation, and comment on the pedagogical value of these implementations. Apart from giving a brief system overview, we report experimental results if available and provide an assessment of how far away the technology is from being deployed in the commercial and educational environments.
A useful and remarkably successful application of speech recognition and processing technology has been demonstrated by a number of research and commercial laboratories in the area of pronunciation training. Voice-interactive pronunciation tutors prompt students to repeat spoken words and phrases or to read aloud sentences in the target language for the purpose of practicing both the sounds and the intonation of the language. The key to teaching pronunciation successfully is corrective feedback, more specifically, a type of feedback that does not rely on the student's own perception. A number of experimental systems have implemented automatic pronunciation scoring as a means to evaluate spoken learner productions in terms of fluency, segmental quality (phonemes) and supra-segmental features (intonation). The automatically generated proficiency score can then be used as a basis for providing other modes of corrective feedback. We discuss segmental and supra-segmental feedback in more detail below.
Segmental Feedback. Technically, designing a voice-interactive pronunciation tutor goes beyond the state of the art required by commercial dictation systems. While the grammar and vocabulary of a pronunciation tutor is comparatively simple, the underlying speech processing technology tends to be complex since it must be customized to recognize and evaluate the disfluent speech of language learners. A conventional speech recognizer is designed to generate the most charitable reading of a speaker's utterance. Acoustic models are generalized so as to accept and recognize correctly a wide range of different accents and pronunciations. A pronunciation tutor, by contrast, must be trained to both recognize and correct subtle deviations from standard native pronunciations.
A number of techniques have been suggested for automatic recognition and scoring of non-native speech (Bernstein, 1997; Franco, Neumeyer, Kim, & Ronen, 1997; Kim, Franco, & Neumeyer, 1997; Witt & Young, 1997). In general terms, the procedure consists of building native pronunciation models and then measuring the non-native responses against the native models. This requires models trained on both native and non-native speech data in the target language, and supplemented by a set of algorithms for measuring acoustic variables that have proven useful in distinguishing native from non-native speech. These variables include response latency, segment duration, inter-word pauses (in phrases), spectral likelihood, and fundamental frequency (F0). Machine scores are calculated from statistics derived from comparing non-native values for these variables to the native models.
In a final step, machine generated pronunciation scores are validated by correlating these scores with the judgment of human expert listeners. As one would expect, the accuracy of scores increases with the duration of the utterance to be evaluated. Stanford Research Institute (SRI) has demonstrated a 0.44 correlation between machine scores and human scores at the phone level. At the sentence level, the machine-human correlation was 0.58, and at the speaker level it was 0.72 for a total of 50 utterances per speaker (Franco et al., 1997; Kim et al., 1997). These results compare with 0.55, 0.65, and 0.80 for phone, utterance, and speaker level correlation between human graders. A study conducted at Entropic shows that based on about 20 to 30 utterances per speaker and on a linear combination of the above techniques, it is possible to obtain machine-human grader correlation levels as high as 0.85 (Bernstein, 1997).
Others have used expert knowledge about systematic pronunciation errors made by L2 adult learners in order to diagnose and correct such errors. One such system is the European Community project SPELL for automated assessment and improvement of foreign language pronunciation (Hiller, Rooney, Vaughan, Eckert, Laver, & Jack, 1994). This system uses advanced speech processing and recognition technologies to assess pronunciation errors by L2 learners of English (French or Italian speakers) and provide immediate corrective feedback. One technique for detecting consonant errors induced by inter-language transfer was to include students' L1 pronunciations into the grammar network. In addition to the English /th/ sound, for example, the grammar network also includes /t/ or /s/, that is, errors typical of non-native Italian speakers of English. This system, although quite simple in the use of ASR technology, can be very effective in diagnosing and correcting known problems of L1 interference. However, it is less effective in detecting rare and more idiosyncratic pronunciation errors. Furthermore, it assumes that the phonetic system of the target language (e.g., English) can be accurately mapped to the learners' native language (e.g., Italian). While this assumption may work well for an Italian learner of English, it certainly does not for a Chinese learner; that is, there are sounds in Chinese that do not resemble any sounds in English.
A system for teaching the pronunciation of Japanese long vowels, the mora nasal, and mora obstruents was recently built at the University of Tokyo. This system enables students to practice phonemic differences in Japanese that are known to present special challenges to L2 learners. It prompts students to pronounce minimal pairs (e.g., long and short vowels) and returns immediate feedback on segment duration. Based on the limited data, the system seems quite effective at this particular task. Learners quickly mastered the relevant duration cues, and the time spent on learning these pronunciation skills was well within the constraints of Japanese L2 curricula (Kawai & Hirose, 1997). However, the study provides no data on long-term effects of using the system.
Supra-segmental Feedback. Correct usage of supra-segmental features such as intonation and stress has been shown to improve the syntactic and semantic intelligibility of spoken language (Crystal, 1981). In spoken conversation, intonation and stress information not only helps listeners to locate phrase boundaries and word emphasis, but also to identify the pragmatic thrust of the utterance (e.g., interrogative vs. declarative). One of the main acoustical correlates of stress and intonation is fundamental frequency (F0); other acoustical characteristics include loudness, duration, and tempo. Most commercial signal processing software have tools for tracking and visually displaying F0 contours (see Figure 2). Such displays can and have been used to provide valuable pronunciation feedback to students. Experiments have shown that a visual F0 display of supra-segmental features combined with audio feedback is more effective than audio feedback alone (de Bot, 1983; James, 1976), especially if the student's F0 contour is displayed along with a native model. The feasibility of this type of visual feedback has been demonstrated by a number of simple prototypes (Abberton & Fourcin, 1975; Anderson-Hsieh, 1994; Hiller et al., 1994; Spaai & Hermes, 1993; Stibbard, 1996). We believe that this technology has a good potential for being incorporated into commercial CALL systems.
Other types of visual pronunciation feedback include the graphical display of a native speaker's face, the vocal tract, spectrum information, and speech waveforms (see Figure 2). Experiments have shown that a visual display of the talker improves not only word identification accuracy (Bernstein & Christian, 1996), but also speech rhythm and timing (Markham & Nagano-Madesen, 1997). A large number of commercial pronunciation tutors on the market today offer this kind of feedback. Yet others have experimented with using a real-time spectrogram or waveform display of speech to provide pronunciation feedback. Molholt (1990) and Manuel (1990) report anecdotal success in using such displays along with guidance on how to interpret the displays to improve the pronunciation of suprasegmental features in L2 learners of English. However, the authors do not provide experimental evidence for the effectiveness of this type of visual feedback. Our own experience with real-time spectrum and waveform displays suggests their potential use as pronunciation feedback provided they are presented along with other types of feedback, as well as with instructions on how to interpret the displays.
Apart from supporting systems for teaching basic pronunciation and literacy skills, ASR technology is being deployed in automated language tutors that offer practice in a variety of higher-level linguistic skills ranging from highly constrained grammar and vocabulary drills to limited conversational skills in simulated real-life situations. Prior to implementing any such system, a choice needs to be made between two fundamentally different system design types: closed response vs. open response design. In both designs, students are prompted for speech input by a combination of written, spoken, or graphical stimuli. However, the designs differ significantly with reference to the type of verbal computer-student interaction they support. In closed response systems, students must choose one response from a limited number of possible responses presented on the screen. Students know exactly what they are allowed to say in response to any given prompt. By contrast, in systems with open response design, the network remains hidden and the student is challenged to generate the appropriate response without any cues from the system.
Closed Response Designs. One of the first implementations of a closed response design was the Voice Interactive Language Instruction System (VILIS) developed at SRI (Bernstein & Rtischev, 1991). This system elicits spoken student responses by presenting queries about graphical displays of maps and charts. Students infer the right answers to a set of multiple-choice questions and produce spoken responses.
A more recent prototype currently under development in SRI is the Voice Interactive Language Training System (VILTS), a system designed to foster speaking and listening skills for beginning through advanced L2 learners of French (Egan, 1996; Neumeyer et al., 1996; Rypa, 1996). The system incorporates authentic, unscripted conversational materials collected from French speakers into an engaging, flexible, and user-centered lesson architecture. The system deploys speech recognition to guide students through the lessons and automatic pronunciation scoring to provide feedback on the fluency of student responses. As far as we know, only the pronunciation scoring aspect of the system has been validated in experimental trials (Neumeyer et al., 1996).
In pedagogically more sophisticated systems, the query-response mode is highly contextualized and presented as part of a simulated conversation with a virtual interlocutor. To stimulate student interest, closed response queries are often presented in the form of games or goal-driven tasks. One commercial system that exploits the full potential of this design is TraciTalk (Courseware Publishing International, Inc., Cupertino, CA), a voice-driven multimedia CALL system aimed at more advanced ESL learners. In a series of loosely connected scenarios, the system engages students in solving a mystery. Prior to each scenario, students are given a task (e.g., eliciting a certain type of information), and they accomplish this task by verbally interacting with characters on the screen. Each voice interaction offers several possible responses, and each spoken response moves the conversation in a slightly different direction. There are many paths through each scenario, and not every path yields the desired information. This motivates students to return to the beginning of the scene and try out a different interrogation strategy. Moreover, TraciTalk features an agent that students can ask for assistance and accepts spoken commands for navigating the system. Apart from being more fun and interesting, games and task-oriented programs implicitly provide positive feedback by giving students the feeling of having solved a problem solely by communicating in the target language.
The speech recognition technology underlying closed response query implementations is very simple, even in the more sophisticated systems. For any given interaction, the task perplexity is low and the vocabulary size is comparatively small. As a result, these systems tend to be very robust. Recognition accuracy rates in the low to upper 90% range can be expected depending on task definition, vocabulary size, and the degree of non-native disfluency.
In the previous sections, we reviewed the current state of speech technology, discussed some of the factors affecting recognition performance, and introduced a number of research prototypes that illustrate the range of speech-enabled CALL applications that are currently technically and pedagogically feasible. With the exception of a few exploratory open response dialog systems, most of these systems are designed to teach and evaluate linguistic form (pronunciation, fluency, vocabulary study, or grammatical structure). This is no coincidence. Formal features can be clearly identified and integrated into a focused task design. This means that robust performance can be expected. Furthermore, mastering linguistic form remains an important component of L2 instruction, despite the emphasis on communication (Holland, 1995). Prolonged, focused practice of a large number of items is still considered an effective means of expanding and reinforcing linguistic competence (Waters, 1994). However, such practice is time consuming. CALL can automate these aspects of language training, thereby freeing up valuable class time that would otherwise be spent on drills.
While such systems are an important step in the right direction, other more complex and ambitious applications are conceivable and no doubt desirable. Imagine a student being able to access the Internet, find the language of his or her choice, and tap into a comprehensive voice-interactive multimedia language program that would provide the equivalent of an entire first year of college instruction. The computer would evaluate the student's proficiency level and design a course of study tailored to his or her needs. Or think of using the same Internet resources and a set of high-level authoring tools to put together a series of virtual encounters surrounding the task of finding an apartment in Berlin. As a minimum, one would hope that natural speech input capacity becomes a routine feature of any CALL application.
To many educators, these may still seem like distant goals, and yet we believe that they are not beyond reach. In what follows, we identify four of the most persistent issues in building speech-enabled language learning applications and suggest how they might be resolved to enable a more widespread commercial implementation of speech technology in CALL.
An intelligent open response language tutor must not only correctly recognize a given speech input, but in addition understand what has been said and evaluate the meaning of the utterance for pragmatic appropriateness. Automatic speech understanding requires Natural Language Processing (NLP) capabilities, a technology for extracting grammatical, semantic, and pragmatic information from written or spoken discourse. NLP has been successfully deployed in expert systems and information retrieval. One of the first voice-interactive dialog systems using NLP was the DARPA-sponsored Air Travel Information System (Pallett, 1995), which enables the user to obtain flight information and make ticket reservations over the telephone. Similar commercial systems have been implemented for automatic retrieval of weather and restaurant information, virtual environments, and telephone auto-attendants. Many of the lessons learned in developing such systems can be valuable for designing CALL applications for practicing conversational skills.