Crowd-Sourced Speech Corpora for Javanese, Sundanese, Sinhala, Nepali, and Bangladeshi Bengali

Oddur Kjartansson, Supheakmungkol Sarin, Knot Pipatsrisawat, Martin Jansche, Linne HaSLTU 2018 · DOI

TL;DR

This paper introduces five new crowd-sourced speech corpora for Javanese, Sundanese, Sinhala, Nepali, and Bangladeshi Bengali, making them publicly available with accompanying Kaldi recipes. This effort aims to democratize speech recognition research for under-resourced languages by providing essential data and tools, thereby lowering the barrier to entry for researchers.

What Problem Does This Paper Address?

Developing speech recognition (ASR) systems for under-resourced languages is a significant challenge due to the scarcity of large, high-quality datasets. Traditional data collection methods are often expensive and time-consuming. This paper addresses this gap by presenting a scalable, crowd-sourced approach to collect speech data for five languages spoken by hundreds of millions of people but historically lacking sufficient digital resources for ASR research.

Methods

The authors employed a crowd-sourcing methodology, leveraging native-speaker volunteers within the respective regions. Recordings were made using consumer electronics in relatively quiet environments. For data collection, they utilized Google's DataHound tool and, for Javanese, also the open-source Eyra tool. The collected data consists of audio recordings and corresponding text transcriptions. To make these corpora usable with the Kaldi ASR toolkit, they developed custom scripts for data conversion and created basic "recipes" that include grapheme-based phonology and lexicons, which are automatically generated from the provided text. These recipes are designed to validate data integrity and demonstrate basic ASR model training rather than achieve state-of-the-art performance.

Datasets

The paper introduces five new speech corpora:

  • Javanese: Collected in Yogyakarta, Indonesia.
  • Sundanese: Collected in Bandung, Indonesia.
  • Sinhala: Collected in Sri Lanka.
  • Bengali: Collected in Bangladesh.
  • Nepali: Collected in Nepal.

Each corpus contains approximately 200,000 recorded utterances. The data is made available under a Creative Commons license. The paper also references existing corpora and tools like Kaldi, Wall Street Journal, and Resource Management for context and comparison.

Key Insights and Ideas

  • Crowd-sourcing as a scalable solution: The paper demonstrates that crowd-sourcing, when managed effectively with local collaborators and volunteers, can be a viable and efficient method for collecting substantial speech data for under-resourced languages.
  • Democratizing ASR research: By releasing these corpora and providing Kaldi recipes, the authors significantly lower the technical and financial barriers for researchers interested in these languages, fostering broader participation in ASR development.
  • Grapheme-based approach for initial validation: The use of automatically generated grapheme-based lexicons and phonology, while not ideal for competitive ASR, serves as a practical and effective method for initial data validation and demonstrating the corpora's utility with standard ASR toolkits.
  • Importance of local collaboration: Partnerships with local universities and institutions were crucial for accessing volunteers and understanding regional nuances, highlighting the value of community engagement in data collection.

Key Results

The paper presents word-error rates (WER) for monophone and triphone models trained on Sinhala, Javanese, and Sundanese data. For example, the triphone model for Sundanese achieved a WER of 3.16%, and for Javanese, it was 9.45%. These results, while "highly optimistic" due to the simplified setup (e.g., no out-of-vocabulary words, language model derived from all data), serve to validate the integrity of the collected data and the effectiveness of the provided Kaldi recipes.

Limitations and Future Work

The authors acknowledge that the grapheme-based approach for lexicons and phonology is a simplification and insufficient for producing competitive ASR systems. They note that for languages with complex writing systems, this naive approach creates a lexical challenge that needs further addressing. Future work includes providing similar recipes for the remaining languages, exploring methodologies to assess data quality more rigorously (similar to references [5] and [7]), and potentially training models jointly across some or all of the collected languages.

Summary updated

BibTeX
@inproceedings{kjartansson2018crowd,
  author = {Oddur Kjartansson and Supheakmungkol Sarin and Knot Pipatsrisawat and Martin Jansche and Linne Ha},
  title = {{Crowd-Sourced Speech Corpora for Javanese, Sundanese, Sinhala, Nepali, and Bangladeshi Bengali}},
  booktitle = {SLTU},
  year = {2018},
  doi = {10.21437/sltu.2018-11},
}
Figure