Lightweight text-to-speech (TTS) models have gained attention for their fast inference and low resource requirements, but they typically depend on external aligners such as the Montreal Forced Aligner (MFA) during training. For Korean, the publicly available MFA acoustic models and pronunciation dictionaries are of limited quality, requiring substantial manual effort to set up a reliable alignment pipeline. In this paper, we propose a lightweight Korean TTS pipeline that can be trained without any external aligner. The proposed method introduces a Connectionist Temporal Classification (CTC) aligner that classifies mel-spectrogram frames directly into phonemes and extracts phoneme durations via forced alignment. Training is organized into two stages: in Stage 1, the CTC aligner is jointly trained with the acoustic model to obtain alignments; in Stage 2, the acoustic model is retrained from scratch using the fixed alignment targets extracted by the Stage 1 aligner. While the monotonic alignment search (MAS) used by larger models fails to converge in lightweight configurations because the encoder representation lacks sufficient capacity, CTC alignment operates directly on the mel spectrogram and is therefore robust to the encoder's capacity. For vocoding, we use a lightweight HiFi-GAN fine-tuned on Korean speech data. On the KSS dataset, the proposed method achieves a MOS of 3.67 without any external aligner, outperforming an MFA-based counterpart (3.46) at the same model scale.
Minhyung Jo (2026) studied this question.