| 1 |
Vocapia-LIMSI System for 2020 Shared Task on Code-switched Spoken Language Identification
|
|
|
|
In: The First Workshop on Speech Technologies for Code-Switching in Multilingual Communities ; https://hal.archives-ouvertes.fr/hal-03091792 ; The First Workshop on Speech Technologies for Code-Switching in Multilingual Communities, Oct 2020, Shanghai, China (2020)
|
|
Abstract:
International audience ; This paper describes the systems submitted by Vocapia Research and LIMSI for the shared task on Code-switched Spoken Language Identification, organized in the conjunction with the First Workshop on Speech Technologies for Code-switching in Multilingual Communities 2020. Our primary system combines an acoustic approach based on i-vector modeling of audio segments with a phonotactic approach that focuses on sequences of language-independent phone units. Both modeling approaches provided comparable performance, and a gain was obtained by a simple linear combination of their scores, showing their complementarity. One of our submissions obtained first rank for all combinations of tasks and language pairs. For the utterancelevel detection task (task A), an F-measure of 76.0% was obtained with our combined system for which the average accuracy on the development set was 83.3%. For the frame-level detection task, the average accuracy was 81.2% on the development set and 78.7% on the evaluation set. However, a detailed analysis reveals a very high rejection of the 200ms codeswitched frames, which comprise only 12% of the corpus. This shows that a more precise modeling of code-switched segments is needed for an accurate segmentation.
|
|
Keyword:
[INFO]Computer Science [cs]; code-switching; language identification; phonotactic model
|
|
URL: https://hal.archives-ouvertes.fr/hal-03091792/file/code-switching-2020.pdf
https://hal.archives-ouvertes.fr/hal-03091792
https://hal.archives-ouvertes.fr/hal-03091792/document
|
|
BASE
|
|
Hide details
|
|
|
|
|
