mwang (02/16/83)
_D_E_P_A_R_T_M_E_N_T _O_F _C_O_M_P_U_T_E_R _S_C_I_E_N_C_E
_U_N_I_V_E_R_S_I_T_Y _O_F _W_A_T_E_R_L_O_O
_S_E_M_I_N_A_R _A_C_T_I_V_I_T_I_E_S
_C_O_M_P_U_T_E_R _S_C_I_E_N_C_E _C_O_L_L_O_Q_U_I_U_M
- Wednesday, February 23, 1983.
Dr. Renato De Mori of the University of Turin, Italy,
will speak on ``_A_l_g_o_r_i_t_h_m_s _a_n_d _A_r_c_h_i_t_e_c_t_u_r_e_s _f_o_r _S_p_e_e_c_h
_U_n_d_e_r_s_t_a_n_d_i_n_g''.
TIME: 3.30 PM
ROOM: MC 5158
ABSTRACT
The purpose of this talk is to present some ideas about
the conception of a speech understanding system as a
distributed knowledge-base system. Unlike other sys-
tems described so far (e.g., HEARSAY-II), a system is
proposed in which lexical and sub-lexical levels are
implemented by cooperative processes interacting as in
a data-flow machine. Such a system organization makes
it possible, for example, to access the lexicon with a
set of phonetic features not rich enough for generating
a lexical hypothesis but capable of focusing the atten-
tion on a set of words compatible with syntactic and
semantic predictions.
In almost all the speech recognition systems proposed
so far the flow of information below lexical level is
only one-way, usually from signal to words. In some
systems, phone symbols, syll-parts and syllables are
generated with a process of template watching followed
by string translation. In the view taken in this talk,
syllable features are not generated by a pure transla-
tion which takes into account only already generated
phone symbols. Rather, generation of syllabic features
may be constrained by lexical expectations and may in-
volve extraction of new information from the data.
February 16, 1983