A four layer sharing HMM system for very large vocabulary isolated word recognition

This paper reports on a large vocabulary speaker independent isolated word recognizer targeting 50,000 words. The system supports a unique four-layer sharing structure for either continuous HMM or discrete HMM. Evaluation is performed using a dictionary of 5000 US city names, a dictionary of the 5000 English most frequent words, a dictionary of 50,000 English words, and the 110,000 word CMU English dictionary. For these dictionaries, recognition accuracy ranges from 90% to 93% for the top 3 results. The speech signal is a one-dimensional waveform as shown in FIGURE 1. The speech signal may be labeled with a sequence of phonemes. A word may correspond to one or more continuous phonemes. An example of the pho-neme labels of the isolated word " item " is also shown in FIGURE 1. FIGURE 1 : Speech waveform and phonemes of the isolated word " item ". A left-to-right HMM process is used to model the speech waveform in this Speech Recognition Engine (SRE) as shown in FIGURE 2. This figure displays a simple 3-state left-to-right HMM of a phoneme where the context includes the left and the right phone, i.e., the HMM is context dependent. This type of HMM is chosen because it offers convenient flexibility for state sharing between the first, second, and last state of the HMM, as explained below. A series of HMMs correspond to a series of pho-nemes. The observations are emitted from each state of the HMM process. The observation probabilities can be formulated as probability distribution , where refers to a state in a HMM. Each state transition, shown as an arc in FIGURE 2, is associated with a state transition probability which denotes the probability of transitioning using the arc of state. FIGURE 2 : A three state left-to-right HMM used in SRE for all the phonemes in a particular context. Suppose there are types of observations, and denotes the distribution of state and type (or stream) , then (1) The observations can be handled in two ways to create either a discrete observation HMM (DHMM) or a continuous observation HMM (CHMM). The probability distribution of the discrete observation HMM defined as (2) is a one-dimension array with each scalar denoting the probability of observing the vector quantized symbol for state , and denoting the sub-probability distribution that is a component in. in the equation denotes the total number of. The sub-probability distribution is introduced …