LSTM-Based Speech Recognition for Assamese: A Low-Resource Language Approach

Technical Summary: LSTM-Based Speech Recognition for Assamese: A Low-Resource Language Approach

Problem Statement
Speech recognition systems have advanced significantly for major languages, yet North Eastern Indian languages, particularly Assamese, remain underexplored. Assamese is characterized as a low-resource language with limited computational linguistic data. Existing research on Assamese speech processing often relies on traditional methods (e.g., Hidden Markov Models) or focuses on limited feature representations, lacking comprehensive analysis at the vowel level. Furthermore, the integration of multiple acoustic features to improve recognition performance in such low-resource contexts remains an open challenge. This study addresses the gap in developing an efficient speech recognition system specifically for Assamese vowels using deep learning techniques.

Methodology
The proposed approach utilizes a Long Short-Term Memory (LSTM) network, a specialized Recurrent Neural Network (RNN) architecture designed to overcome the vanishing gradient problem and learn long-term temporal dependencies in sequential data.

• Dataset: The authors curated a self-developed dataset comprising 1,760 vowel samples. Data was collected from 22 speakers (14 male, 8 female) from the Assamese community. Each of the eight distinct Assamese vowels was recorded 10 times per speaker. The dataset was split into training (70%), testing (15%), and validation (15%) sets.
• Feature Extraction: To enhance speech representation, the study integrates three distinct acoustic features:
  1. Mel Frequency Cepstral Coefficients (MFCC): 13 coefficients (12 MFCC + 1 energy) based on the human auditory system's non-linear pitch perception.
  2. Short-Term Energy (STE): Measures the energy present in a specific time segment of the signal.
  3. Zero Crossing Rate (ZCR): Counts the number of times the signal changes sign within a frame, distinguishing between voiced and unvoiced signals.
     These features were combined into a 15-dimensional feature vector (1 STE + 1 ZCR + 13 MFCC).
• Model Architecture: The LSTM model consists of an input layer, a hidden layer with 64 LSTM units, and a fully connected output layer with units corresponding to the eight vowel classes. The output layer utilizes a SoftMax activation function.
• Training Configuration: The model was trained for 60 epochs with a batch size of 32. The Categorical Cross-Entropy loss function was used, optimized via the Adam optimizer with a learning rate of 0.001.

Key Contributions
The paper outlines five primary contributions:

1. Development of a self-curated, balanced Assamese vowel speech dataset.
2. Integration of MFCC, STE, and ZCR features to create a robust speech representation.
3. Design and implementation of an LSTM-based recognition model tailored for the Assamese language.
4. Comparative evaluation against existing deep learning models (BLSTM and SincConv+DSConv).
5. Analysis of performance degradation caused by phonetic similarities among specific Assamese vowels.

Results
The system was evaluated using Precision, Recall, F1-Score, Accuracy, and Word Error Rate (WER).

• Overall Performance: The model achieved an average classification accuracy of 95%, with an average Precision of 84%, Recall of 83%, and F1-Score of 83%. The average Word Error Rate (WER) was 18.60%.
• Per-Class Analysis: High accuracy (99%) was observed for vowels /aa/ and /ea/. However, performance degraded for vowels /u/, /o/, and /a'/ (represented as 'oo' and 'ow' in the text), showing lower Precision, Recall, and F1-Scores.
• Error Analysis: The authors attribute the higher misclassification rates and elevated WER for specific vowels to their phonetic similarities. Spectrograms of /u/, /a'/, and /o/ revealed nearly similar formant movements, leading to confusion in the model and increased substitution errors.
• Comparative Performance: When compared to existing models (BLSTM and SincConv+DSConv) using the same feature set, the proposed LSTM model demonstrated superior performance, achieving 95% accuracy compared to 93.4% and 92% respectively, and a lower WER (18.60% vs. 19.40% and 20.30%).

Significance and Claims
The authors position this work as a foundational step toward preserving and developing Assamese language technologies. By demonstrating that a combination of standard acoustic features and LSTM networks can achieve high accuracy (95%) even with a relatively small, self-curated dataset, the study validates the feasibility of deep learning for low-resource languages.

The paper modestly acknowledges its limitations, noting that the small dataset size may affect generalization and that the current scope is limited to isolated vowel recognition. The authors emphasize that while the classification accuracy is high, the WER remains a challenge due to inherent phonetic ambiguities in the language. They conclude that this work provides a baseline for future improvements, suggesting that larger datasets, advanced architectures, and optimization techniques could further enhance recognition rates for continuous speech and keyword detection.