Extending the Hierarchical Scheme: Inter and Intra Phonetic Information - Hierarchical Neural Network Structures for Phoneme Recognition - page 98

Fig. 5.25. Combination of inter and intra-phonetic schemes. The intra phonetic
scheme is implemented by the STC system, where the temporal context has been
split in two blocks. Taken from [Schwarz 08] page 36.
The STC scheme forms a hierarchical MLP structure consisting of two
MLPs situated at the first hierarchical level and one MLP situated at the
second level. According to our work, this scheme can be classified as an intra-
phonetic scheme since the three MLPs involved in the system are trained
based on the same label, corresponding to the central label of the entire
context. In addition, the MLPs of the first hierarchical level are specialized
in different parts of the entire context.
The first row of Table 5.17 shows PA of the STC system 2 where each MLP
has 500 hidden units 3 . Following the proposed approach, the intra and inter-
phonetic schemes are combined when the intra-phonetic approach is imple-
mented by the STC system as shown in Fig. 5.25, analogous to the serial
combinationtechnique shownin Fig. 5.19. Table 5.17 shows results when a win-
dow downsampling is applied at the input of the inter-phonetic scheme.
These experiments show again the considerable advantage of combining
inter and intra-phonetic information. This approach achieves a maximum
2 The phonetic recognizer has been taken from the STC software package
(http://speech.fit.vutbr.cz/en/software/phoneme-recognizer-based-long-
temporal-context).
3 This result slightly differs from published one (75,59%) due to implementation in
the evaluation process, e.g. Viterbi decoder tool and the exact value of phoneme
insertion penalty.