Transliteration-Based Approaches for Multilingual and Code-Switched Languages

Explore a comprehensive lecture on transliteration-based approaches for multilingual and code-switched languages. Delve into the challenges of evaluating code-mixed language ASR systems and learn about the innovative transliteration-optimized Word Error Rate (toWER) metric. Discover a novel language-agnostic multilingual ASR system that utilizes a many-to-one transliteration transducer to map similar acoustics to a single canonical target sequence. Examine the benefits of this approach for Indic languages, achieving up to 10% relative reduction in Word Error Rate compared to language-dependent models. Gain insights from Bhuvana Ramabhadran, a distinguished researcher in multilingual speech recognition and synthesis, as she shares her expertise on addressing the complexities of code-switching and multilingual ASR systems.