• Dartmouth LING 48 Final Project: Improving TTS for Shanghainese
• Yuanhao Chen [email protected] Spring 2023

To build a text-to-speech (TTS) system for Shanghainese from scratch, seeking to improve the production of tone sandhi compared to existing models by paying special attention to preprocessing of text.

See writeup/main.pdf.

pip install -r phonemisation/requirements.txt
pip install -r speech_synthesis/requirements.txt
pip install -r comparison_questionnaire/requirements.txt  # for analysis of questionnaire results

See speech_synthesis/README.md.

• phonemisation/: contains the phonemisation module
  • See explanation of output in phonemisation/__init__.py
  • Usage: python -m phonemisation "text to phonemise"
  • Mechanism: Chinese sentence — word segmentation ⟶ Chinese words — romanisation ⟶ Shanghainese pinyin — phonemisation ⟶ Shanghainese phonemes
    • jieba is used for word segmentation
    • A Shanghainese dictionary I previously made is used for romanisation
      • Uses Qieyun module to add the tone number 1 to syllables of 陰平 yinping/inbin tone; other tones are phonologically unmarked
    • The romanisation_to_ipa function in romanisation.py contains the phonemisation function
• make_metadata.py: uses the phonemisation module to convert transcription into IPA and generate metadata for training
  • See below in data/
• data/: contains the dataset used for training
  • The transcriptions and audio files are adapted from this repo
    • Downsampled to 16kHz for training
    • Currently, only shh.dict.cn/ is used for training
  • The */metadata.txt files are generated by make_metadata.py
• training/
  • Juptyer notebook for training the model
  • Intended to be uploaded and run in Google Colab environment; needs to be modified for local use
  • Uses the coqui-ai/TTS repo, which contains an implementation of VITS
• writeup/: the write-up
• speech_synthesis/: contains the speech synthesis model
  • See speech_synthesis/README.md for more details
• comparison_questionnaire/: contains the questionnaire and audio files used to compare speech produced by this model, the Apple model, and a human speaker
  • *-1.wav: produced by this model
  • *-2.wav: produced by Apple VoiceOver (MacBook Pro 14-inch, 2021; MacOS Ventura 13.0.1)
  • *-3.wav: spoken by myself
  • stats.ipynb: Jupyter notebook for analysing the questionnaire results