HiACC: Hinglish adult & children code-switched corpus

Area of Science:

• Computational Linguistics
• Speech Processing
• Sociolinguistics

Background:

• Code-switching, the alternation between languages, is common among bilinguals, particularly in India with Hinglish (Hindi-English).
• Existing automatic speech recognition (ASR) systems struggle with code-switched data due to a lack of diverse, representative datasets, leading to higher word error rates (WER).
• Current ASR models show a 30-50% increase in WER when processing code-switched speech compared to monolingual speech.

Purpose of the Study:

• To introduce the HiACC corpus, a benchmark Hinglish speech dataset designed to enhance ASR performance in resource-constrained environments.
• To address the scarcity of publicly available code-switched speech datasets, especially those including children's speech.
• To provide a valuable resource for linguistic and computational analysis of code-switching.

Main Methods:

• Development of the HiACC corpus, comprising 3,318 audio segments from adults and 1,858 from children.
• Inclusion of 5.24 hours of read and spontaneous Hinglish speech with detailed annotations and code-switching tags.
• Public release of the corpus with segmented audio and aligned transcripts for open research.

Main Results:

• Baseline ASR experiments demonstrated that models trained on monolingual data underperform significantly, with a WER of approximately 42% on the Hinglish test set.
• The HiACC corpus provides the first publicly available resource for code-switched Hinglish speech that includes both adult and child speakers.
• The dataset's availability is expected to catalyze progress in developing more robust ASR systems for code-switched languages.

Conclusions:

• The HiACC corpus is a crucial resource for advancing research in automatic speech recognition for code-switched languages like Hinglish.
• The findings highlight the challenges ASR systems face with code-switched input and the need for specialized datasets.
• This work paves the way for more inclusive and accurate speech recognition technologies for diverse linguistic communities.