Mircea Giurgiu * Results on Automatic Speech Recognition in Romanian

3. Experiments for isolated word recognition

3.1. Speech recognition results using Dynamic Time Warping technique

Since speaking rate variations cause nonlinear fluctuations in speech pattern time axis, the elimination of these fluctuations has been accomplished by DTW using dynamic programming. The word with the minimum distance is selected as the recognized word. This is time-expensive even on powerful computers, so we have proposed the idea to quantize the features extracted from isolated words. In this manner each spectral feature is vector quantified using a VQ dictionary and the distances between the centroids vectors are precomputed in the training stage. So, what we need when a distance between two vectors has to be computed is to point in a look-up table and get the appropriate distance. In this case, different experiments to evaluate the recognition performances on Romanian spoken digit recognition have been done: monolocutor and multilocutor with or without VQ and for different VQ sizes [3]. Here is reported only the multilocutor experiment, when the test speech database was VQ-coded with variable size (Table 4).

Table 4. The recognition percentage for VQDTW multilocutor experiment

Word	Recognized (M=32)	Nonrecognized (M=32)	Percent (%)	Recognized(M=64)	Nonrecognized (M=64)	Percent (%)
"Unu"	15	0	100	15	0	100
"Doi"	13	2	86.66	13	2	86.66
"Trei"	15	0	100	14	1	93.33
"Patru"	15	0	100	15	0	100
"Cinci"	15	0	100	15	0	100
"ªase"	13	2	86.66	14	1	93.33
"ªapte"	13	2	86.66	14	1	93.33
"Opt"	11	4	73.33	12	3	80
"Nouã"	15	0	100	15	0	100
"Zece"	15	0	100	15	0	100
Total	140	10	93.33	142	8	94.66

From the VQDTW-type of experiments we have concluded that the recognition performance depends on the VQ size M (32 and 64) being 94%, respectively 96%, for monolocutor experiments and 93.33% and 94.66% for multilocutor experiments. Even though the recognition rate is poorer for VQDTW than for DTW, the advantages consist in the memory size and computing speed, so a possibility to implement the algorithms on DSP platforms arouses.

3.2. Hidden Markov Models applied for automatic recognition of Romanian digits

A HMM is a model of speech generation, although it is used for speech recognition purposes. It is characterized by five elements: (a) A state set S with N states S={s₁,s₂,...,s_N}. Denote the current state at time t as q_t. (b) A symbol set with M different symbols V={v₁,...,v_M}. Current symbol at t, O_t. (c) A transition probability distribution A={a_ij}, where a_ij is the transition probability from state s_i to state s_j. (d) A symbol probability distribution B={b_j(k)}, where b_j(k) is the production probability of symbol v_k at state s_j. (e) An initial state distribution (probability that the symbol generation begins in state s_i).

For a vocabulary with L words {W₁, W₂,...,W_L}, the recognition system is a set with L HMM , where each model is associated with one word W_i of the vocabulary. Given an input symbol sequence, the generation probability is computed for every model . The model giving the highest probability corresponds to the recognized word. For speech recognition, it is suitable to use the left-to-right model, because it matches the sequential nature of the speech generation process [8].

The vocabulary consists of the ten Romanian digits recorded under normal conditions of work room with computer noise at 10 kHz and 16 bits/sample. LPC analysis has been applied and then VQ with the codebook size M=64. We have recorded a large number of speakers (40) uttering 3 times each word, but this communication presents only two subsets of this large database. One is for training the HMM and is composed of 3 speakers, and the other contains the testing subjects, totally independent from the first one. Ten HMM have been designed for each word. During the training step, the number of states (N) of HMM was variable, in order to detect which is the best number of states. Theoretically, the larger N is, the better the results should be. This is not always true, and an N greater than 6 does not much improve the performances (Fig. 10). Since the Romanian speech database was too small, the HMM were verified with a Spanish database obtained from Granada University. The results were similar and they demonstrated the validity of our results [3,8].

Fig. 10. - The recognition error versus the number of states in HMM.

182