Fabian Santiago

Combining Continuous Word Representation and Prosodic Features for ASR Error Prediction

Recent advances in continuous word representation have been successfully used in several natural language processing tasks. This paper focuses on error prediction in Automatic Speech Recognition ASR outputs and proposes to investigate the use of continuous word representation word embeddings within a neural network architecture. The main contribution of this paper is about word embeddings combination: several combination approaches are proposed in order to take advantage of their complementarity. The use of prosodic features, in addition to classical syntactic ones, is evaluated. Experiments are made on automatic transcriptions generated by the LIUM ASR system applied on the ETAPE corpus. They show that the proposed neural architecture, using an effective continuous word representation combination and prosodic features as additional features, outperforms significantly state-of-the-art approach based on the use of Conditional Random Fields. Last, the proposed system produces a well calibrated confidence measure, evaluated in terms of Normalized Cross Entropy.

The Acquisition of Question Intonation by Mexican Spanish Learners of French

In this chapter, we analyze the final tunes and the prosodic structure observed in yes/no and wh-questions in French as an L2 produced by Mexican Spanish learners. Our study consists in a cross-comparison of information-seeking interrogatives recorded in French and Mexican Spanish in various settings and produced by 15 Mexican learners of French (L2), 10 French speakers, and 10 Mexican speakers. Analyses of the data show some differences between native and nonnative productions: (i) an overuse of rising tunes is observed in the learners’ productions in all question types; in addition, an extra-rising contour is frequently used; and (ii) the internal prosodic structure in long sentences is generally not marked by tonal cues (e.g., pitch accents) in learners’ productions. These patterns could be partly viewed as resulting from an L1 transfer in the case of yes/no questions, since similar prosodic patterns were found in the Spanish native speakers’ productions. However, this hypothesis is not confirmed by the analysis of wh-questions: native speakers have a tendency to use a large variety of final tunes whereas learners use almost only rising contours, in particular, the extra-rising one. These results could lead to consider that L1 transfer in the acquisition of L2 prosody does not account for all learners’ prosodic patterns. Alternate hypotheses may be put forward to account for the realizations observed: (i) a prosodic simplification, or (ii) the idea that some rises may be used to express some sort of linguistic insecurity.

The role of lexical stress on vowel duration and vowel space in two varieties of Spanish

This paper investigates the effects of lexical stress on vowel durations, vowel space and vowel quality in Spanish. Data come from oral productions of 22 Spanish speakers (10 from Madrid and 12 from Mexico City) performing different tasks. As for durational cues, we found that vowel durations play a role as a cue of lexical stress. Interestingly, our results also show differences between the two varieties (the stressed-unstressed ratio being larger for Mexican than Madrilenian speakers). Instead, we show that the expansion/compression of the vowel space is not affected by lexical stress, but it does seem to be affected by the task type. We found, however, that lexical stress can affect vowel quality in certain cases: unstressed /a/ and /o/ tend to be centralized. We discuss these results in the light of previous research reporting effects of lexical stress on vowel spectral quality in different varieties of Spanish