26-31 May 2013

ICASSP

Google

Cross-lingual

1. Feature learning: In this approach, features extraction is learnt independent of the language/s. Therefore the same steps are used for any future languages. Afterwards an ASR system can be built on top of these learnt (from data) features. One major drawback is the system is no optimised jointly.

   [Cross-lingual and Multi-stream Posterior Features for Low Resource LVCSR Systems](https://jointphd.notion.site/Cross-lingual-and-Multi-stream-Posterior-Features-for-Low-Resource-LVCSR-Systems-b8ca6e7c29ab4a53a08d7f7d6f89422f)

2. Transfer learning: In this approach, a model learnable weights are initialised from a pre-trained (resource rich) model.

   1. Pre-training can also be done using unsupervised learning and more recently using self supervised learning.

Multilingual

1. Multi-task learning: Training with all the languages combined.

trains acoustic model based on DNN.

Dataset: The experiments are performed on the Romance languages including Catalan, different Spanish dialects, French, Italian, two Portuguese dialects, Romanian, and Basque (which is not a Romance language but a language isolate surrounded by Romance languages), see Table 1.