No 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 ...

This paper describes speech signal modeling techniques which are well suited to high performance and robust isolated word recognition. We present new techniques for incorporating spectral/temporal information as a function of temporal position within each word. In particular, spectral/temporal parameters are computed using both variable length blocks with a variable spacing between blocks. We tested features computed with these methods using an alphabet recognition task based on the ISOLET database. The Hidden Markov Model Toolkit (HTK) was used to implement the isolated word recognizer with whole word HMM models. The best accuracy achieved for speaker independent alphabet recognition, using 50 features, was 97.9%, which represents a new benchmark for this task. We also tested these methods with deliberate signal degradation using additive Gaussian noise and telephone band limiting and found that the recognition degrades gracefully and to a smaller degree than for control cases based on MFCC coefficients and delta cepstra terms.

No abstract

No abstract

No abstract

Music selection is difficult without efficient organization based on metadata or tags, and one effective tag scheme is based on the emotion expressed by the music. However, manual annotation is labor intensive and unstable because the perception of music emotion varies from person to person. This paper presents an emotion classification system for digital music with a resolution of eight emotional classes. Russell’s emotion model was adopted as common ground for emotional annotation. The music information retrieval (MIR) toolbox was employed to extract acoustic features from audio files. The classification system utilized a supervised machine learning technique to recognize acoustic features and create predictive models. Four predictive models were proposed and compared. The models were composed by crossmatching two types of neural networks, i.e., Levenberg-Marquardt (LM) and resilient backpropagation (Rprop), with two types of structures: a traditional multiclass model and the cascaded structure of a binary-class model. The performance of each model was evaluated via the MediaEval Database for Emotional Analysis (DEAM) benchmark. The best result was achieved by the model trained with the cascaded Rprop neural network (accuracy of 89.5%). In addition, correlation coefficient analysis showed that timbre features were the most impactful for prediction. Our work offers an opportunity for a competitive advantage in music classification because only a few music providers currently tag music with emotional terms.

No abstract