Table 2 Frobenius distance and R2 values between the phonetic similarity matrix (Fig. 5C) and log-probability matrices (Fig. 5A) for each of the four NN models.

From: Wordsworth: A generative word dataset for comparison of speech representations in humans and neural networks

 

Frobenius distance (a.u.)

R2 (a.u.)

Recurrent modified 2D cochleagram

10.3 [12.4, 12.8]

10.7 [12.3, 12.7]

0.14 [2.4 × 10−7, 1.5 × 10−3]

0.10 [2.6 × 10−7, 1.4 × 10−3]

Modified 2D cochleagram

11.2 [13.3, 13.8]

11.9 [13.5, 13.9]

0.12 [2.8 × 10−7, 1.5 × 10−3]

0.06 [3.2 × 10−7, 1.5 × 10−3]

2D cochleagram48

13.0 [14.4, 14.7]

13.9 [14.9, 15.2]

0.08 [3.4 × 10−7, 1.5 × 10−3]

0.06 [3.0 × 10−7, 1.4 × 10−3]

1D waveform42

10.3 [12.3, 12.7]

11.3 [12.6, 13.0]

0.13 [3.2 × 10−7, 1.5 × 10−3]

0.07 [2.8 × 10−7, 1.5 × 10−3]

  1. Considering that the phonetic similarity matrix is symmetric and the log-probability matrices are asymmetric, we report both measures for both upper and lower triangles of the log-probability matrices. Numbers in brackets represent the 95% confidence intervals for the corresponding null distributions. All true Frobenius distance and R2 measures: p < 0.0001.
Back to article page