CubicMonk19/phonemeza

Live grapheme-to-phoneme tool for isiZulu, isiXhosa & Afrikaans; a from-scratch PyTorch attention seq2seq, served via FastAPI, deployed on AWS. # PhonemeZA — South African grapheme-to-phoneme PhonemeZA predicts the pronunciation of isiZulu, isiXhosa, and Afrikaans words from their spelling, using a sequence-to-sequence model whose LSTM cells and attention are implemented from scratch in PyTorch (no `nn.LSTM`). Type a word and it returns the phoneme sequence in X-SAMPA plus a heatmap of which input letters the decoder attended to at each output step. **▶ Live demo: phonemeza.duckdns.org** ## Results Test-set metrics (held-out 10% split; PER = phone error rate, the mean edit distance between predicted and reference phoneme sequences normalised by reference length): | Language | Context | PER | Word accuracy | |----------|---------|----:|--------------:| | isiZulu | attention | 0.0031 | 98.7% | | isiXhosa | attention | 0.0017 | 99.0% | | Afrikaans | attention | 0.0284 | 82.4% | | Afrikaans | bottleneck (no attention) | 0.0894 | 69.6% | The Nguni languages (isiZulu, isiXhosa) score near-perfect, but that number deserves caveats. Their orthographies are close to phonemic, the NCHLT Southern-Bantu dictionaries are largely rule-derived, and the languages are agglutinative — so stems and affixes recur across the train/test boundary even though the word lists are disjoint. Afrikaans, whose pronunciations are less predictable from spelling, is the truer test, and there the attention decoder cuts PER roughly 3× and raises word accuracy from 69.6% to 82.4% over a bottleneck decoder that must compress the whole word into one fixed vector. That replicates the classic attention-vs-bottleneck finding from the original CMUdict experiments on a new set of languages. ## Architecture - **From-scratch recurrence.** `LSTMCell` and an attention-augmented `LSTMCellWithContext` are written out gate-by-gate (`g2p/model.py`); the encoder and decoder stack these manually rather than calling `nn.LSTM`. - **Dot-product attention.** At each decoding step the decoder scores its hidden state against every encoder state, …