next up previous contents index
Next: 1.3 Signal Representation Up: Spoken Language Input Previous: 1.1 Overview

1.2: Speech Recognition

Victor Zue, Ron Cole, & Wayne Ward
MIT Laboratory for Computer Science, Cambridge, Massachusetts, USA
Oregon Graduate Institute of Science & Technology, Portland, Oregon, USA
Carnegie Mellon University, Pittsburgh, Pennsylvania, USA

Defining the Problem

Speech recognition is the process of converting an acoustic signal, captured by a microphone or a telephone, to a set of words. The recognized words can be the final results, as for applications such as commands & control, data entry, and document preparation. They can also serve as the input to further linguistic processing in order to achieve speech understanding, a subject covered in section gif.

Speech recognition systems can be characterized by many parameters, some of the more important of which are shown in Figure gif. An isolated-word speech recognition system requires that the speaker pause briefly between words, whereas a continuous speech recognition system does not. Spontaneous, or extemporaneously generated, speech contains disfluencies, and is much more difficult to recognize than speech read from script. Some systems require speaker enrollment---a user must provide samples of his or her speech before using them, whereas other systems are said to be speaker-independent, in that no enrollment is necessary. Some of the other parameters depend on the specific task. Recognition is generally more difficult when vocabularies are large or have many similar-sounding words. When speech is produced in a sequence of words, language models or artificial grammars are used to restrict the combination of words.

The simplest language model can be specified as a finite-state network, where the permissible words following each word are given explicitly. More general language models approximating natural language are specified in terms of a context-sensitive grammar.

One popular measure of the difficulty of the task, combining the vocabulary size and the language model, is perplexity, loosely defined as the geometric mean of the number of words that can follow a word after the language model has been applied (see section gif for a discussion of language modeling in general and perplexity in particular). Finally, there are some external parameters that can affect speech recognition system performance, including the characteristics of the environmental noise and the type and the placement of the microphone.

Table: Typical parameters used to characterize the capability of speech recognition systems

Speech recognition is a difficult problem, largely because of the many sources of variability associated with the signal. First, the acoustic realizations of phonemes, the smallest sound units of which words are composed, are highly dependent on the context in which they appear. These phonetic variabilities are exemplified by the acoustic differences of the phonemegif /t/ in two, true, and butter in American English. At word boundaries, contextual variations can be quite dramatic---making gas shortage sound like gash shortage in American English, and devo andare sound like devandare in Italian.

Second, acoustic variabilities can result from changes in the environment as well as in the position and characteristics of the transducer. Third, within-speaker variabilities can result from changes in the speaker's physical and emotional state, speaking rate, or voice quality. Finally, differences in sociolinguistic background, dialect, and vocal tract size and shape can contribute to across-speaker variabilities.

Figure gif shows the major components of a typical speech recognition system. The digitized speech signal is first transformed into a set of useful measurements or features at a fixed rate, typically once every 10--20 msec (see sections gif and 11.3 for signal representation and digital signal processing, respectively). These measurements are then used to search for the most likely word candidate, making use of constraints imposed by the acoustic, lexical, and language models. Throughout this process, training data are used to determine the values of the model parameters.

Figure: Components of a typical speech recognition system.

Speech recognition systems attempt to model the sources of variability described above in several ways. At the level of signal representation, researchers have developed representations that emphasize perceptually important speaker-independent features of the signal, and de-emphasize speaker-dependent characteristics [Her90]. At the acoustic phonetic level, speaker variability is typically modeled using statistical techniques applied to large amounts of data. Speaker adaptation algorithms have also been developed that adapt speaker-independent acoustic models to those of the current speaker during system use, (see section gif). Effects of linguistic context at the acoustic phonetic level are typically handled by training separate models for phonemes in different contexts; this is called context dependent acoustic modeling.

Word level variability can be handled by allowing alternate pronunciations of words in representations known as pronunciation networks. Common alternate pronunciations of words, as well as effects of dialect and accent are handled by allowing search algorithms to find alternate paths of phonemes through these networks. Statistical language models, based on estimates of the frequency of occurrence of word sequences, are often used to guide the search through the most probable sequence of words.

The dominant recognition paradigm in the past fifteen years is known as hidden Markov models (HMM). An HMM is a doubly stochastic model, in which the generation of the underlying phoneme string and the frame-by-frame, surface acoustic realizations are both represented probabilistically as Markov processes, as discussed in sections gif, gif and 11.2. Neural networks have also been used to estimate the frame based scores; these scores are then integrated into HMM-based system architectures, in what has come to be known as hybrid systems, as described in section 11.5.

An interesting feature of frame-based HMM systems is that speech segments are identified during the search process, rather than explicitly. An alternate approach is to first identify speech segments, then classify the segments and use the segment scores to recognize words. This approach has produced competitive recognition performance in several tasks [ZGPS90,FBC95].

1.2.2 State of the Art

Comments about the state-of-the-art need to be made in the context of specific applications which reflect the constraints on the task. Moreover, different technologies are sometimes appropriate for different tasks. For example, when the vocabulary is small, the entire word can be modeled as a single unit. Such an approach is not practical for large vocabularies, where word models must be built up from subword units.

Performance of speech recognition systems is typically described in terms of word error rate, E, defined as:

where N is the total number of words in the test set, and S, I, and D are the total number of substitutions, insertions, and deletions, respectively.

The past decade has witnessed significant progress in speech recognition technology. Word error rates continue to drop by a factor of 2 every two years. Substantial progress has been made in the basic technology, leading to the lowering of barriers to speaker independence, continuous speech, and large vocabularies. There are several factors that have contributed to this rapid progress. First, there is the coming of age of the HMM. HMM is powerful in that, with the availability of training data, the parameters of the model can be trained automatically to give optimal performance.

Second, much effort has gone into the development of large speech corpora for system development, training, and testing. Some of these corpora are designed for acoustic phonetic research, while others are highly task specific. Nowadays, it is not uncommon to have tens of thousands of sentences available for system training and testing. These corpora permit researchers to quantify the acoustic cues important for phonetic contrasts and to determine parameters of the recognizers in a statistically meaningful way. While many of these corpora (e.g., TIMIT, RM, ATIS, and WSJ; see section 12.3) were originally collected under the sponsorship of the U.S. Defense Advanced Research Projects Agency (ARPA) to spur human language technology development among its contractors, they have nevertheless gained world-wide acceptance (e.g., in Canada, France, Germany, Japan, and the U.K.) as standards on which to evaluate speech recognition.

Third, progress has been brought about by the establishment of standards for performance evaluation. Only a decade ago, researchers trained and tested their systems using locally collected data, and had not been very careful in delineating training and testing sets. As a result, it was very difficult to compare performance across systems, and a system'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 computer technology have also indirectly influenced our progress. The availability of fast computers 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 high-end workstations without additional hardware---a 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, speaker-independent recognition of digit strings spoken continuously and restricted to telephone bandwidth can achieve an error rate of 0.3% when the string length is known.

One of the best known moderate-perplexity tasks is the 1,000-word so-called Resource Management (RM) task, in which inquiries can be made concerning various naval vessels in the Pacific ocean. The best speaker-independent performance on the RM task is less than 4%, using a word-pair 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 isolated-word, speaker-dependent systems for many years, the community has since 1992 moved towards very-large-vocabulary (20,000 words and more), high-perplexity (), speaker-independent, continuous speech recognition. The best system in 1994 achieved an error rate of 7.2% on read sentences drawn from North America business news [PFF94].

With the steady improvements in speech recognition performance, systems are now being deployed within telephone and cellular networks in many countries. Within the next few years, speech recognition will be pervasive in telephone networks around the world. There are tremendous forces driving the development of the technology; in many countries, touch tone penetration is low, and voice is the only option for controlling automated services. In voice dialing, for example, users can dial 10--20 telephone numbers by voice (e.g., call home) after having enrolled their voices by saying the words associated with telephone numbers. AT&T, on the other hand, has installed a call routing system using speaker-independent word-spotting technology that can detect a few key phrases (e.g., person to person, calling card) in sentences such as: I want to charge it to my calling card.

At present, several very large vocabulary dictation systems are available for document generation. These systems generally require speakers to pause between words. Their performance can be further enhanced if one can apply constraints of the specific domain such as dictating medical reports.

Even though much progress is being made, machines are a long way from recognizing conversational speech. Word recognition rates on telephone conversations in the Switchboard corpus are around 50% [CGF94]. It will be many years before unlimited vocabulary, speaker-independent continuous dictation capability is realized.

1.2.3 Future Directions

In 1992, the U.S. National Science Foundation sponsored a workshop to identify the key research challenges in the area of human language technology, and the infrastructure needed to support the work. The key research challenges are summarized in [CH92]. Research in the following areas for speech recognition were identified:


In a robust system, performance degrades gracefully (rather than catastrophically) as conditions become more different from those under which it was trained. Differences in channel characteristics and acoustic environment should receive particular attention.


Portability refers to the goal of rapidly designing, developing and deploying systems for new applications. At present, systems tend to suffer significant degradation when moved to a new task. In order to return to peak performance, they must be trained on examples specific to the new task, which is time consuming and expensive.


How can systems continuously adapt to changing conditions (new speakers, microphone, task, etc) and improve through use? Such adaptation can occur at many levels in systems, subword models, word pronunciations, language models, etc.


How can systems continuously adapt to changing conditions (new speakers, microphone, task, etc) and improve through use? Such adaptation can occur at many levels in systems, subword models, word pronunciations, language models, etc.

Language Modeling:

Current systems use statistical language models to help reduce the search space and resolve acoustic ambiguity. As vocabulary size grows and other constraints are relaxed to create more habitable systems, it will be increasingly important to get as much constraint as possible from language models; perhaps incorporating syntactic and semantic constraints that cannot be captured by purely statistical models.

Confidence Measures:

Most speech recognition systems assign scores to hypotheses for the purpose of rank ordering them. These scores do not provide a good indication of whether a hypothesis is correct or not, just that it is better than the other hypotheses. As we move to tasks that require actions, we need better methods to evaluate the absolute correctness of hypotheses.

Out-of-Vocabulary Words:

Systems are designed for use with a particular set of words, but system users may not know exactly which words are in the system vocabulary. This leads to a certain percentage of out-of-vocabulary words in natural conditions. Systems must have some method of detecting such out-of-vocabulary words, or they will end up mapping a word from the vocabulary onto the unknown word, causing an error.

Spontaneous Speech:

Systems that are deployed for real use must deal with a variety of spontaneous speech phenomena, such as filled pauses, false starts, hesitations, ungrammatical constructions and other common behaviors not found in read speech. Development on the ATIS task has resulted in progress in this area, but much work remains to be done.


Prosody refers to acoustic structure that extends over several segments or words. Stress, intonation, and rhythm convey important information for word recognition and the user's intentions (e.g., sarcasm, anger). Current systems do not capture prosodic structure. How to integrate prosodic information into the recognition architecture is a critical question that has not yet been answered.

Modeling Dynamics:

Systems assume a sequence of input frames which are treated as if they were independent. But it is known that perceptual cues for words and phonemes require the integration of features that reflect the movements of the articulators, which are dynamic in nature. How to model dynamics and incorporate this information into recognition systems is an unsolved problem.

next up previous contents
Next: 1.3 Signal Representation Up: Spoken Language Input Previous: 1.1 Overview