Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Phonetically rich corpus construction for a low-resourced language

Marcellus Amadeus, William Alberto Cruz Castañeda, Wilmer Lobato, Niasche Aquino

TL;DR

This work tackles the challenge of creating a phonetically rich corpus for Brazilian Portuguese, a low-resource language, to improve acoustic modeling and TTS/ASR applications. It introduces a four-stage methodology—linguistic-variable establishment, diverse corpus collection, textual processing, and recording protocols—built around a phoneme-to-triphone framework with acoustic-articulatory (vocoid/contoid) classification. The approach yields a corpus (Alana AI) with substantially higher distinct triphone coverage than CETUC, TTS-Portuguese, and Globo, achieving a $55.8\%$ increase in distinct triphones at comparable sizes and saturation around $5{,}000$ sentences when maximizing linguistic diversity. The work provides a concrete, replicable workflow for constructing phonetically rich data in low-resource settings, with clear implications for improving the naturalness and accuracy of PT-BR speech technologies across ASR and TTS tasks.

Abstract

Speech technologies rely on capturing a speaker's voice variability while obtaining comprehensive language information. Textual prompts and sentence selection methods have been proposed in the literature to comprise such adequate phonetic data, referred to as a phonetically rich \textit{corpus}. However, they are still insufficient for acoustic modeling, especially critical for languages with limited resources. Hence, this paper proposes a novel approach and outlines the methodological aspects required to create a \textit{corpus} with broad phonetic coverage for a low-resourced language, Brazilian Portuguese. Our methodology includes text dataset collection up to a sentence selection algorithm based on triphone distribution. Furthermore, we propose a new phonemic classification according to acoustic-articulatory speech features since the absolute number of distinct triphones, or low-probability triphones, does not guarantee an adequate representation of every possible combination. Using our algorithm, we achieve a 55.8\% higher percentage of distinct triphones -- for samples of similar size -- while the currently available phonetic-rich corpus, CETUC and TTS-Portuguese, 12.6\% and 12.3\% in comparison to a non-phonetically rich dataset.

Phonetically rich corpus construction for a low-resourced language

TL;DR

This work tackles the challenge of creating a phonetically rich corpus for Brazilian Portuguese, a low-resource language, to improve acoustic modeling and TTS/ASR applications. It introduces a four-stage methodology—linguistic-variable establishment, diverse corpus collection, textual processing, and recording protocols—built around a phoneme-to-triphone framework with acoustic-articulatory (vocoid/contoid) classification. The approach yields a corpus (Alana AI) with substantially higher distinct triphone coverage than CETUC, TTS-Portuguese, and Globo, achieving a increase in distinct triphones at comparable sizes and saturation around sentences when maximizing linguistic diversity. The work provides a concrete, replicable workflow for constructing phonetically rich data in low-resource settings, with clear implications for improving the naturalness and accuracy of PT-BR speech technologies across ASR and TTS tasks.

Abstract

Speech technologies rely on capturing a speaker's voice variability while obtaining comprehensive language information. Textual prompts and sentence selection methods have been proposed in the literature to comprise such adequate phonetic data, referred to as a phonetically rich \textit{corpus}. However, they are still insufficient for acoustic modeling, especially critical for languages with limited resources. Hence, this paper proposes a novel approach and outlines the methodological aspects required to create a \textit{corpus} with broad phonetic coverage for a low-resourced language, Brazilian Portuguese. Our methodology includes text dataset collection up to a sentence selection algorithm based on triphone distribution. Furthermore, we propose a new phonemic classification according to acoustic-articulatory speech features since the absolute number of distinct triphones, or low-probability triphones, does not guarantee an adequate representation of every possible combination. Using our algorithm, we achieve a 55.8\% higher percentage of distinct triphones -- for samples of similar size -- while the currently available phonetic-rich corpus, CETUC and TTS-Portuguese, 12.6\% and 12.3\% in comparison to a non-phonetically rich dataset.
Paper Structure (14 sections, 5 figures, 6 tables)

This paper contains 14 sections, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Initial guidelines
  3. Methodology
  4. Establishing linguistic variables
  5. Textual corpora selection
  6. Exclusion criteria
  7. Textual processing
  8. Corpora size selection
  9. Sentences selection
  10. Protocols: speakers and recordings
  11. Discussion and results
  12. Conclusion
  13. Acknowledgements
  14. Bibliographical References

Figures (5)

  • Figure 1: Methodology for building a phonetically rich text corpus for PT-BR.
  • Figure 2: Phonemic histogram of the CETUC corpus text by considering a vocoid-contoid triphone classification.
  • Figure 3: Phonemic histogram of the CETUC corpus text by considering a vocoid-contoid triphone classification.
  • Figure 4: Changepoint analysis for the variance of new triphones per sentence.
  • Figure 5: Corpora new triphones per sentence.