Signal modeling for isolated word recognition

Abstract

This paper presents speech signal modeling techniques which are well suited to high performance and robust isolated word recognition. Speech is encoded by a discrete cosine transform of its spectra, after several preprocessing steps. Temporal information is then also explicitly encoded into the feature set. We present a new technique for incorporating this temporal information as a function of temporal position within each word. We tested features computed with this method using an alphabet recognition task based on the ISOLET database. The HTK toolkit was used to implement the isolated word recognizer with whole word HMM models. The best result obtained based on 50 features and speaker independent alphabet recognition was 98.0%. Gaussian noise was added to the original speech to simulate a noisy environment. We achieved a recognition accuracy of 95.8% at a SNR of 15 dB. We also tested our recognizer with simulated telephone quality speech by adding noise and band limiting the original speech. For this "telephone" speech, our recognizer achieved 89.6% recognition accuracy. The recognizer was also tested in a speaker dependent mode, resulting in 97.4% accuracy on test data. INTRODUCTIONContinuous speech recognition systems have been developed for many real-world applications, often using commercial low-cost speech recognition software. However, high performance and robust isolated word recognition, particularly for the letters of the alphabet recognizer and for digits, is still useful for many applications such as recognizing telephone numbers, spelled names and address, and ZIP codes.Because of the potential applications, as mentioned above, many isolated word recognizers are optimized for the digits or alphabet or both (alphadigit). The alphabet recognition task is particularly difficult because there are many highly confusable letters in the alphabet set---for example the great acoustic similarity among the letters of the E-set (b, c, d, e, g, p, t, v, z) or for the (m,n) pair. Also, since language models cannot generally be used, the alphabet recognition task is a small, challenging, and potentially useful problem for evaluating acoustic signal modeling and word recognition methodsSeveral techniques have been proposed to improve isolated word recognition systems. For example, the best result in a speaker independent alphabet recognition was obtained using a multi-tier phoneme-based Hidden Markov Model (HMM) recognizer [5]. Disadvantages of phoneme-based HMM recognizers are the system complexity and the phonetic transcription of the training words has to be known.The main contribution of this paper is to present a method for isolated word recognition which is easier to implement than the state of the art systems introduced to date, and one which gives better performance than any of these previously introduced systems.The ISOLET database, [1], was used for all experiments reported in this paper. This LDC distributed database was intended for evaluation of isolated word recognizers and it has therefore ...

Signal modeling for high-performance robust isolated word recognition

“…The block features were recomputed every 10 ms. No manual segmentation or phonetic labeling was required or used. The primary modification, relative to [8] and [9], is that the block length was varied at both ends of each analyzed utterance, rather than only for the beginning section. See Fig.1 for an illustration of this variable block length method.…”

Nonlinear Dimensionality Reduction Methods for Use with Automatic Speech Recognition

“…For both training and testing data, the modified Discrete Cosine Transformation Coefficients (DCTC) and Discrete Cosine Series Coefficients (DCSC) (Zahorian et al 1991;Zahorian et al, 1997;Zahorian et al, 2002;Karnjanadecha & Zahorian, 1999) were extracted as original features. The modified DCTC is used for representing speech spectra, and the modified DCSC is used to represent spectral trajectories.…”

Spoken English Alphabet Recognition with Mel Frequency Cepstral Coefficients and Back Propagation Neural Networks

“…In conclusion, we can say that some of the disadvantage of phoneme based recognizers as in [1] when compared to word based recognizer is complexity of the system and the word transcription must be known [2]. Also, from our review, many of these testing and experiments were done using HMMs and modified techniques.…”

“…Signal modeling for high performance and robust isolated word recognition were proposed in [2,8]. The authors proposed a new technique for incorporating temporal and spectral feature within each word.…”

“…Isolated speech recognition system may also use spoken digits to test its recognition accuracy. Spoken alphabet recognition may have several applications among them, automated directory assistance to retrieve information such as spelling names, telephone numbers addresses and ZIP codes [1,2].Spoken alphabet recognition may be seen as a simple task for human beings but unfortunately, for machines this can be a challenging task due to high acoustic similarities among certain groups of letters [1][2][3]. High acoustic similarities may cause difficulty in classification while low acoustic similarities causes ease to discriminate among classes for speech recognition systems.An alphabet set which has been identified to be the most confusable for speech recognition is the so called E-set letters.…”