[Paper Review] Comprehending Real Numbers: Development of Bengali Real Number Speech Corpus

Md Mahadi Hasan Nahid, Md. Ashraful Islam+2 more|arXiv (Cornell University)|Mar 27, 2018

Speech Recognition and SynthesisComputer Science5 references4 citations

TL;DR

This paper introduces the first comprehensive Bengali real number speech corpus, comprising 2,302 audio samples from 10 male speakers across 3.79 hours, covering all possible Bengali real number words and phrases. It enables training and evaluation of Bengali automatic speech recognition systems, achieving a 15% word error rate with CMU Sphinx4 and 13.2% with an LSTM-based recognizer, highlighting the need for balanced, diverse corpora in low-resource languages.

ABSTRACT

Speech recognition has received a less attention in Bengali literature due to the lack of a comprehensive dataset. In this paper, we describe the development process of the first comprehensive Bengali speech dataset on real numbers. It comprehends all the possible words that may arise in uttering any Bengali real number. The corpus has ten speakers from the different regions of Bengali native people. It comprises of more than two thousands of speech samples in a total duration of closed to four hours. We also provide a deep analysis of our corpus, highlight some of the notable features of it, and finally evaluate the performances of two of the notable Bengali speech recognizers on it.

Motivation & Objective

• To address the lack of comprehensive Bengali speech datasets for real numbers, which hinders automatic speech recognition (ASR) development.
• To create a balanced, representative corpus covering all possible Bengali real number utterances, including numbers from 0 to 100 and multi-digit structures like 'hundred', 'thousand', 'lakh', 'crore', and 'decimal'.
• To provide a benchmark for evaluating Bengali ASR systems using both traditional HMM-based and deep learning-based models.
• To analyze phonetic and lexical characteristics of the corpus, including phoneme frequency imbalances and word length distributions.
• To lay the foundation for future expansion with more diverse speakers and broader linguistic coverage.

Proposed method

• Generated synthetic speech utterances using a structured algorithm that randomly combines base number words (0–99), hundred-scale words (100–900), and magnitude terms (thousand, lakh, crore, decimal).
• Ensured balanced word frequency distribution across the 115-word vocabulary to support even model training, though prior probabilities were not modeled due to lack of authentic data.
• Collected speech samples from 10 native Bengali speakers (ages 20–23) across different regions, storing them as .wav files with unique identifiers and corresponding transcripts in text-data.txt.
• Organized the corpus into speaker-specific directories and used standardized naming (e.g., speaker_5_21.wav) for reproducibility and data management.
• Applied two distinct ASR models: CMU Sphinx4 (HMM-based) and a bidirectional LSTM with post-processing, both evaluated using word and phoneme error rates.
• Conducted error rate analysis across varying training durations and evaluated performance on both training and held-out test sets (80:20 split).

Experimental results

Research questions

• RQ1 How can a comprehensive and balanced Bengali real number speech corpus be systematically constructed to cover all possible number utterances in the language?
• RQ2 What are the key phonetic and lexical characteristics of Bengali real number speech, including phoneme frequency imbalances and word length distributions?
• RQ3 How do traditional HMM-based and modern deep learning-based ASR models perform on this newly created Bengali speech corpus?
• RQ4 To what extent does training data size affect the word error rate in Bengali ASR systems, particularly for number recognition?
• RQ5 What are the limitations of the current corpus in terms of speaker diversity and phonetic coverage, and how can they be addressed in future work?

Key findings

• The corpus contains 2,302 speech samples, 17,582 total words, and spans approximately 3.79 hours of audio, with 115 unique vocabulary items covering all real number expressions in Bengali.
• The corpus includes 102 base number words (0–100) and 13 special terms (e.g., 'hajar' for thousand, 'koTi' for crore), with 45 having two possible pronunciations.
• The CMU Sphinx4 model achieved a 15% word error rate on a held-out 20% test set, demonstrating baseline performance on the new corpus.
• The bidirectional LSTM model achieved a 13.2% word error rate and 29% phoneme error rate, with performance improving over time due to post-processing.
• Significant phoneme frequency imbalances were observed—e.g., 'A' had 10.8% normalized frequency, while 'AI' had only 0.15%—indicating inherent linguistic disparities in the language.
• The corpus lacks female speakers and has limited age diversity, suggesting a need for future expansion to improve model generalization and representativeness.