chaumi-k/carnatic-bilstm-transcription

Carnatic BiLSTM+CTC Swara Transcription

Audio → CQT → 3× BiLSTM(256) + CTC → swara sequence. Trained on the CompMusic Carnatic Varnam dataset (28 recordings, 7 ragas). Labels follow the dataset's own svara convention. Evaluated leave-one-raga-out. See config.json for feature params, normalization, vocab, and raga maps. Reference: Saputra et al. 2021 (BiLSTM AMT backbone).