Synthetic speech augmentation for low-resource Estonian and Slovenian ASR: Comparing Parakeet-TDT and Whisper

Estonian and Slovenian are morphologically rich languages with few transcribed hours available for training. In both, a single edit can change case or number while leaving a well-formed word, so an aggregate error rate cannot distinguish a meaning-altering error from a trivial one. We adapt two pretrained multilingual models on identical data for each language: the 0.6B-parameter Parakeet-TDT transducer and the 1.55B-parameter Whisper-large-v3 encoder–decoder. Each model is trained on CommonVoice 17 alone and on CommonVoice 17 augmented with synthetic speech generated from language-model text. Zero-shot inference together with four training configurations forms a controlled ladder of adaptation states, so the augmentation experiment functions as the measuring instrument for the subsequent error analysis rather than as an independent result. Synthetic-speech augmentation improves in-domain recognition for every architecture-language pair, and under each model’s standard full fine-tuning recipe, Parakeet, the smaller model, attains the lower word error rate in every fine-tuned condition. We analyze error signatures over 95,070 re-aligned sentence pairs, supported by an exhaustive manual reading of a 1735-utterance stratified sample, to separate model-specific error patterns from adaptation effects. Whisper’s excess of complete misrecognitions holds in every fine-tuned regime of both languages and is already present zero-shot in Estonian, and Parakeet’s language-conditioned compound-boundary bias holds in every adaptation state, whereas Whisper’s out-of-domain degradation on Slovenian is largest after real-data-only fine-tuning and is only partially recovered by synthetic speech. These results indicate that, across the regimes tested, system choice largely determines which error types persist, while the adaptation regime mainly changes their volume. We release the fine-tuned models, the synthetic corpora, and the per-sentence analysis artifacts.