1: Classification of 8 languages along rhythmic dimensions %V and ∆C. Figure reproduced from Ramus, Nespor, and Mehler (1999).

Contexts in source publication

Context 1

... samples were thus manually segmented into CV sequences, and for each utterance three different measures capturing the proportion of V segments (%V ) and the standard deviation of V and C segments (∆V & ∆C) were computed. Their results showed that %V combined with ∆C resulted in an apparent clustering of the languages (shown in Figure 2.1) into rhythmic classes. ...

Context 2

... this section we will briefly describe the pipeline of the i-vector model (shown in Figure 2.2). For a detailed description of each step, see the Appendix of the present chapter (Section 2.A). ...

Context 3

... -English 60% Figure 2.3: PCA visualization of utterances produced by 4 bilingual speakers from the UCAM corpus. ...

Context 4

... question is of great importance to hypothesize whether infants would be able to easily group their bilingual input by language and not by speaker. In Figure 2.4 we visualize the first two PCA dimensions of the i-vectors of all 4 language pairs. ...

Context 5

... in previous experiments, we first computed a PCA on the i-vectors for each language pair in order to visualize the data. Figure 2.6 shows the first two dimensions of the PCA for each pair. As can be seen in the visualization, all language pairs appear to show a higher degree of overlap compared to the unfiltered speech, suggesting that the model was relying to some extent on phonological differences. ...

Context 6

... results are shown in Table 2.3. Figure 2.6: PCA visualization of low-pass filtered utterances produced by 4 bilingual speakers from the UCAM corpus. ...

Context 7

... in , for each language pair available in the corpus, we simulated an experiment with 40 subjects (20 control and 20 experimental). An example of a simulated experiment is given in Figure 2.7a. interpreted the p-value of control vs. experimental conditions as an indicator of the discriminability of each language pair, with p < 0.05 indicating a significant difference between both groups (thus implying discrimination of the language pair), and p ≥ 0.05 meaning no significant group difference was found (i.e., no language discrimination). ...

Context 8

... the p-value of control vs. experimental conditions as an indicator of the discriminability of each language pair, with p < 0.05 indicating a significant difference between both groups (thus implying discrimination of the language pair), and p ≥ 0.05 meaning no significant group difference was found (i.e., no language discrimination). When running our replication, we first observed that repeating the simulation of 40 subjects several times sometimes changed the interpretation of the discriminability for certain pairs, with p-values going from significant to non-significant (see an example of a stable language pair, Dutch-Japanese, and a language pair with variable p-values, Dutch-English, in Figure 2.7b). These unstable results may be due to the large number of possible utterance sets, and to the fact that we did not control them for duration and syllable length. ...

Context 9

... we repeated the simulations using unfiltered speech. Figure 2.9 shows the matrix of results. ...

Context 10

... computed the ABX score for each of the 4 language pairs using the i-vectors extracted from the four new models. Figure 2.11 shows a comparison of all the models, including the one from Experiment 2. As can be seen, the results obtained for the three French background models are very similar, indicating that the models have learnt similar features from their input despite having been exposed to different speakers. ...

Context 11

... speaker-based LDA rescues the mixed background models from being utterly confused. Finally, we applied several clustering methods to the test i-vectors. For each of the three background conditions, we computed clustering purity as a function of the number of clusters in each level of the hierarchical tree using eight different linkage methods. Fig. 2 shows the results before and after LDA for a complete linkage clustering 1 from K = 1 to K = ...

Context 12

... Figure 2.12, we show a visualization of the old post-LDA utterances (from Experiment 6) and the new utterances, projected onto the same LDA dimensions learnt in Experiment 6. ...

Context 13

... many different sets of features can be used for this purpose, here we will use the popular MFCC-SDC features which, as discussed before, can capture both static and dynamic information. For every step in the feature extraction process (shown in Figure 2.13) we will discuss, when possible, the parallels between the algorithms and human auditory perception. ...

Context 14

... resulting MFCCs are D-dimensional feature vectors that capture short-term spectral information that is relevant for speech recognition. By repeating this process over every frame along the signal, any given utterance can be translated into an MFCC sequence or matrix, as shown in Figure 2.19. ...

Context 15

... algorithm is initialized by setting arbitrary weights, means and covariances ω 0 k , µ 0 k , Σ 0 k for each component (see Figure 2.22, Step 0). ...

Context 16

... whose children learnt French from a nanny or at daycare), and finally only 2 had one parent who spoke both languages equally often. We then examined differences in the way parents used the languages with the child and with other people. Figure 3.2 shows a histogram of the languages used by fathers and mothers when addressing their infants (Fig. 3.2a) and when talking to other people in front of the child (Fig. 3.2b) as reported in the LEQ. As can be seen in the figure on the left, in spite of generally adhering to an OPOL division of languages, some parents reported also speaking a small amount of the other language when talking to their child (cases mostly-FR, mostly-L2 ). This ...

Context 17

... finally only 2 had one parent who spoke both languages equally often. We then examined differences in the way parents used the languages with the child and with other people. Figure 3.2 shows a histogram of the languages used by fathers and mothers when addressing their infants (Fig. 3.2a) and when talking to other people in front of the child (Fig. 3.2b) as reported in the LEQ. As can be seen in the figure on the left, in spite of generally adhering to an OPOL division of languages, some parents reported also speaking a small amount of the other language when talking to their child (cases mostly-FR, mostly-L2 ). This behavior was more common in L2 speakers than in French speakers, ...

Context 18

... responses were automatically collected by the app and coded as a binary dependent variable representing whether the child clicked on the picture or the cross in each trial. Figure 2 shows the proportion of trials where the picture of the familiar object was chosen, split by condition. Responses were analyzed with a generalized linear mixed model (GLMM) with binomial family and logit link using package lme4 Table 2. ...