Hierarchical Tamil Phoneme Classification Using Support Vector Machines

1. Introduction

Tamil, one of the longestsurviving classical languages, has a rich phonetic inventory that includes vowels, consonants, and a set of diacriticdriven allophones. Automatic speech recognition (ASR) and languagetechnology tools for Tamil are still lagging behind those for IndoEuropean languages, partly because of a scarcity of robust phonemelevel classifiers.

This page describes a hierarchical approach to Tamil phoneme classification that leverages the discriminative power of Support Vector Machines (SVM). By arranging phonemes into logical groups (vowels, stops, nasals, etc.) and training a cascade of binary and multiclass SVMs, the system achieves higher accuracy while keeping computational complexity modest.

2. Background

2.1 Tamil Phoneme Set

Tamil consists of 12 short vowels, 5 long vowels, 18 consonants (including 5 retroflexes), and a set of compound letters formed by combining vowels with consonants. For practical classification, the inventory is usually reduced to 3035 distinct phoneme categories. Table1 shows a typical reduction.

2.2 Support Vector Machines

SVMs are maximummargin classifiers that can handle highdimensional data efficiently. Two kernel choices dominate speech tasks:

• Linear kernel suitable when the feature space already separates the classes.
• Radial basis function (RBF) kernel provides nonlinear decision boundaries without explicit feature mapping.

The hierarchical model uses a linear kernel for the toplevel binary decisions (e.g., vowel vs. consonant) and an RBF kernel for deeper, more ambiguous splits (e.g., retroflex vs. nonretroflex).

3. Methodology

3.1 Data Acquisition and Preprocessing

We employed the IIITMadras Tamil Speech Corpus, containing 25hours of read speech from 150 speakers (balanced gender and age). The audio was segmented into phonemelevel units using forced alignment with a preexisting GMMHMM model.

Preprocessing steps:

1. Resampling to 16kHz, 16bit PCM.
2. Preemphasis filter (=0.97).
3. Framing (25ms windows, 10ms shift) and Hamming windowing.

3.2 Feature Extraction

The classifier relies on a combination of spectral and temporal descriptors:

• 13 MelFrequency Cepstral Coefficients (MFCC) + + (total 39 dimensions).
• 13 Linear Predictive Cepstral Coefficients (LPCC).
• Pitch, energy, and zerocrossing rate.
• Formant frequencies (F1, F2, F3) extracted via LPC.

All features are normalized to zero mean and unit variance on a perspeaker basis.

3.3 Hierarchical Architecture

The hierarchy consists of three levels:

1. Level1 Broad class separation: vowel vs. consonant (binary SVM).
2. Level2 Consonant subgrouping: stop vs. nasal vs. fricative/approximant vs. retroflex (fourclass SVM).
3. Level3 Finegrained phoneme identification: dedicated onevsrest SVMs for each phoneme within its subgroup.

Each node outputs a confidence score; the path with the highest cumulative confidence is selected as the final phoneme label.

3.4 Training Procedure

for level in hierarchy:    X_train, y_train = load_features(level)    if level == 1:        kernel = 'linear'    else:        kernel = 'rbf'    svm = SVC(C=1.0, kernel=kernel, probability=True)    svm.fit(X_train, y_train)    save_model(svm, level)

Hyperparameters (C, ) were tuned using 5fold crossvalidation on the training split (70% of the data). The best configuration was C=2.0, =0.01 for the RBF kernels.

3.5 Evaluation Metrics

We report:

• Overall phoneme accuracy (PA)
• Classwise precision, recall, and F1score
• Confusion matrix visualisation

Statistical significance was assessed with McNemars test against a flatmulticlass SVM baseline.

4. Experimental Results

4.1 Baseline Comparison

The flat multiclass SVM (singlelevel, 30 classes) achieved 78.4% phoneme accuracy on the heldout test set. The hierarchical model raised accuracy to 86.9% an improvement of **8.5percentage points**.

4.2 Classwise Performance

4.3 Confusion Analysis

The most common errors involved confusions between the retroflex and the alveolar r, as well as between aspirated stops kh and unaspirated k. These errors were reduced after adding pitchcontour features, indicating that suprasegmental cues help separate phonetically similar categories.

4.4 Ablation Study

Removing Level2 (i.e., directly feeding Level1 outputs to Level3) reduced overall accuracy to 82.1%, confirming the importance of an intermediate subgrouping step.

4.5 Computational Considerations

Training time per level:

• Level1: ~3minutes (linear kernel)
• Level2: ~7minutes (RBF, 4class)
• Level3: ~25minutes (30 onevsrest SVMs, parallelised)

Inference latency averaged 2.3ms per phoneme on a standard laptop CPU, making the model suitable for realtime ASR frontends.

5. Conclusion

The hierarchical SVM framework presented here offers a practical balance between accuracy and efficiency for Tamil phoneme classification. By exploiting the natural linguistic grouping of phonemes, the system outperforms a flat multiclass baseline while keeping training and inference costs modest.

Future work will explore:

• Integration with deep neural embeddings (e.g., xvectors) as additional features.
• Extension to diphone and triphone contexts for improved coarticulation modelling.
• Transfer learning from larger IndoAryan speech corpora to further boost lowresource performance.

Ultimately, a robust phoneme recogniser is a cornerstone for building highquality Tamil speechtotext, voiceassistant, and languagelearning applications.