next up previous contents
Next: Conclusion Up: Integration Models of Previous: General principles for

Five models of audio-visual integration in speech perception

In the area of speech perception, several models have been proposed to account for speech recognition in noise or for the perception of paradoxical stimuli [52,208,209,53,292,348]. However, Summerfield addressed the main theoretical issues and discussed them in the light of experimental data and of general theories of speech perception [332]. He described five different architectural structures that may solve the problem of audio-visual fusion in the perception of speech. We here revisit those five models in order to better fit them into the general principles presented in the above section. The proposed structures are presented in the following sections. They are finally put into relation with the general principles.

Direct Identification Model (DI)
The Direct Identification model of audio-visual configurations is based upon the Lexical Access From Spectra by Klatt [157]. It has been turned into a Lexical Access from Spectra and Face Parameters. The input signals are here directly transmitted to the bimodal classifier. This classifier may be, for instance, a bimodal (or bivectorial) lexicon in which the prototype the closest to the input is selected (see figure C.1 ).

Basics of the model: There is no common representation level over the two modalities between the signal and the percept.

  
Figure C.1 : Schematic of the Direct Identification (DI) model

Separated Identification Model (SI)
In the Separated Identification model, the visual and the auditory input are separately identified through two parallel identification processes. Fusion of the phonemes or phonetic features across each modality occurs after this identification. Fusion can be processed with logical data, such as in the VPAM model (Vision Place Auditory Mode) where each modality deals with a set of phonetic features [220,332]. Thus, the place of articulation of the output is that of the visual input and the mode of articulation (voiced, nasal, etc.) is that of the auditory input (see figure C.2 ).

  
Figure C.2 : Schematic of the VPAM model (example of an SI structure) (after [332])

Fusion can also be processed with probabilistic data (fuzzy logic). Each input can be matched against (unimodal) prototypes in order to get two scores for each category. Then, each pair of data corresponding to the same category is fused through probabilistic computation. The FLMP (Fuzzy Logical Model of Perception) falls into this class of models (see [208,52]).

Basics of the model: The inputs are matched against prototypes (or even classified) before being integrated.

Dominant Modality Recoding Model (RD)
The model of Recoding into the Dominant Modality sees the auditory modality as a dominant modality in the perception of speech since it is better suited to it. The visual input is thus recoded into a representation of this dominant modality. The selected representation of the auditory modality is the transfer function of the vocal tract. This transfer function is separately evaluated from the auditory input (e.g., through cepstral processing) and from the visual input (through association). Both evaluations are then fused. The source characteristics (voiced, nasal, etc.) are evaluated only from the auditory information. The so evaluated pair source / vocal tract is then sent to the phonemic classifier (see figure C.3 ).

  
Figure C.3 : Schematic of the Dominant Modality Recoding (RD) model

Basics of the model: The visual modality is recoded into an auditory metric. Both representations are then integrated into the auditory space where the categorization takes place.

Motor Space Recoding Model (MR)
In the Motor Space Recoding Model, the two inputs are projected onto an amodal space (which is neither auditory nor visual). They are then fused within this common space. This amodal space is the articulatory space of vocal tract configurations. The visual modality is projected only on the vocal tract space dimensions where to information can be carried. For instance, the visual modality may bring information on lip rounding, not on the velum position. The final representation depends on two possibilities. when there is a single projection on a dimension (from the auditory input), this projection makes the final decision. When information from the two modalities is simultaneously projected upon one dimension (e.g. jaw height), the final value comes from a weighted sum of both inputs. The final representation thus obtained is given by the phonemic classifier. This architecture fits well the Motor Theory of Speech Perception [187]) and the Realistic Direct Theory of Speech Perception [108].

  
Figure C.4 : Schematic of the Motor Space Recoding (MR) model

Basics of the model: The two modalities are projected upon a common motor space where they are integrated before final categorization.

Model of Motor Dynamics
This last model is the one least developed by Summerfield. Each modality gives an estimation of the motor dynamics (mass and force) of the vocal tract. Such an estimation might be based upon kinematics data (position, velocity and acceleration). The two pieces of estimation (from each modality) of motor dynamics are fused and then categorized, describing the identity of phonemes in terms of time-varying functions rather than articulatory states traversed [332, page 44,]. The important difference between this model and the previous one relies in the nature of the motor representation, as seen as the projection and integration of the sensory inputs. In fact, the previous model could be given the name AR (Articulatory Recoding) whereas the current model could be entitled MR in itself.

However, introducing such a distinction which is relevant to deal with theories of speech perception is somewhat disputable when dealing with architectures of audiovisual fusion. We could thus think of other distinctions that make use of the various propositions found in literature on the nature of the preprocessing at the input or of the level of the linguistic units at the output. This is why we prefer to take only in consideration the architecture called MR, as described in figure C.4 , since it can regroup ``static'' as well as ``dynamic'' representations of the motor configurations.



next up previous contents
Next: Conclusion Up: Integration Models of Previous: General principles for



Esprit Project 8579/MIAMI (Schomaker et al., '95)
Thu May 18 16:00:17 MET DST 1995