This project implements a Grapheme-to-Phoneme (G2P) conversion system using a sequence-to-sequence model with attention. The model is trained on Romanian grapheme-phoneme pairs, utilizing both phoneme error rate (PER) and word error rate (WER) as evaluation metrics.
Below is a visualization of the training and validation loss over time:
- Phoneme Error Rate (PER): 0.728%
- Word Error Rate (WER): 8.905%
The model demonstrates a steady decrease in both training and validation loss, with the validation loss stabilizing at a low value, indicating effective generalization.
The encoder processes the grapheme sequence and generates a context-aware representation of the input. It uses:
- Bidirectional RNN: Captures both forward and backward dependencies in the sequence.
- Input Size: Determined by the dimensionality of the grapheme input.
The decoder generates the phoneme sequence conditioned on the encoder’s outputs. It incorporates:
- Attention Mechanism: Computes a weighted context vector based on encoder outputs and the decoder’s current state.
- Teacher Forcing: Balances between ground truth phoneme inputs and predicted phonemes during training.
Attention helps the decoder focus on relevant parts of the input sequence during each decoding step.
- Hidden Size: Scaled for bidirectional input to double the dimensionality.
- Learning Rate: Dynamically decayed during training for stabilization.
- Loss Function: Cross-entropy loss computed on phoneme predictions.
The training process includes:
- Dynamic Learning Rate Adjustment: The learning rate is halved if validation loss plateaus for more than 4 consecutive steps.
- Early Stopping: Training stops if the learning rate falls below
1e-5. - Batch Size: Mini-batch gradient descent with a fixed size for efficient training.
Here are the results for test samples:
| Romanian Word | Predicted Phonemes |
|---|---|
| bună | b u n a t |
| dimineață | d i m i n ea ts at |
| copac | k o p a k |
| pădure | p at d u r e |
| cer | tS e r |
| coloreze | k o l o r e z e |
| soare | s oa r e |
| lună | l u n at |
| stea | s t ea |
| floare | f l oa r e |
- Strong Generalization: Low validation loss and error rates indicate the model generalizes well on unseen data.
- Effective Attention Mechanism: The attention mechanism enables the model to focus on relevant graphemes during decoding, improving accuracy.
- Training Challenges: Slight overfitting during early epochs was mitigated with dropout and learning rate adjustments.