Korean SpeechT5 (Jamo Tokenizer, KSS)

@misc{ahnhs2k_speecht5_korean,  
  author = {Ahn, Hosung},  
  title = {Korean SpeechT5 TTS Model},  
  year = {2025},  
  publisher = {Hugging Face},  
  url = {https://huggingface.co/ahnhs2k/...}  
}

모델 특징

• Base Model: microsoft/speecht5_tts
• Dataset: Bingsu/KSS_Dataset
• Tokenizer: Jamo-based Korean tokenizer (character-level, placeholder aware)
• Speaker Embedding: microsoft/wavlm-base-plus-sv
• Vocoder: microsoft/speecht5_hifigan
• Sample Rate: 16 kHz
• 단일 화자 한국어 TTS 모델

Text Utils (korean_text_utils.py)

• number_to_korean(): 숫자 → 한국어 읽기
• normalize_korean(): 추론용 텍스트 정규화
• prosody_split(): 구두점 기반 segment 분할
• inject_tokens_for_training(): 훈련용 placeholder 전처리
• decompose_jamo_with_placeholders(): placeholder 보존 자모 분해

라이선스

license: apache-2.0
datasets:
- Bingsu/KSS_Dataset
language:
- ko
base_model:
- microsoft/speecht5_tts
pipeline_tag: text-to-speech