I'm Junseok Oh (오준석) aka june-oh / RiceBerry

Fully frozen Whisper + Gemma with lightweight adapter-only bridging (0.44% params). 26.8% WER reduction on academic domains via inference-time domain prompting. PAKDD 2026 Accepted (Oral).

Encoder-decoder with variable-rate generation via cross-attention. Frozen speech encoder + LLM LoRA. 2.6%/5.2% WER on LibriSpeech, 4.7% on cross-domain TED-LIUM-v2. EACL 2026 Findings Accepted.

End-to-end SVC pipeline: collected ~10h speaker data, extracted vocals via UVR5, trained so-vits-svc & whisper-vits-svc on RTX A5000. Performed voice conversion and singing voice conversion inference.

Developed universal Korean ASR using hybrid FastConformer RNNT+CTC. Implemented cache-aware streaming for low-latency inference and context biasing for gaming domain vocabulary.

Developed streaming/non-streaming Korean ASR pipelines optimized for 8 kHz telephony using FastConformer-CTC. Implemented dynamic context biasing for domain-shift word-level accuracy.

Wav2Vec multi-task framework for L2-Korean assessment: jointly models pronunciation, fluency, and content. Combined Conformer-CTC ASR with LLaMa for multi-aspect automated scoring.

AI framework for dysarthria severity classification with multi-modal explanations. Implemented speech-based explainable diagnostic modules analyzing acoustic and linguistic features.

Designed sound event detection models for automated video content rating. Fine-tuned Whisper ASR on domain-specific video corpora for robust transcription in diverse acoustic environments.

Built L2-Korean evaluation pipeline combining Conformer-CTC ASR with BERT-based semantic scoring. Developed algorithms to quantify pronunciation accuracy, speech rate, and syntactic correctness.

Modified Kaldi sentence-level decoder for sub-1.0 RT real-time video QA. Collected domain-specific audio/text corpora, retrained acoustic/language models for improved QA accuracy on complex narratives.