Please enable JavaScript.

Coggle requires JavaScript to display documents.

TTS, Novel NLP Methods for Improved Text-To-Speech Synthesis 2021, Novel…

- - - - Part-of-speech tagging
        
        Identifying the grammatical role of each word (e.g., noun, verb).
      - Tokenization
        
        Splitting text into words and punctuation.
  - - - Pronunciation lexicon
        
        A dictionary of words and their pronunciations
      - Rules-based approach
        
        Using linguistic rules to predict pronunciation
      - Statistical approach
        
        Using machine learning to predict pronunciation based on patterns in training data
      - Hybrid approach
  - - - Pitch contour
      - Duration
      - Emotion
      - Pauses
  - - - Convert phonemes to Mel-spectrogram
    - - translates the mel-spectrom into an actual audio waveform
        
        Natural vocoders
        
        WaveNet, WaveRNN, or Parallel WaveGAN
        
        Griffin-Lim algorithm
        
        Traditional method for waveform reconstruction from spectrograms.
    - - Concatenative synthesis
      - Parametric synthesis
      - Articulatory synthesis
    - - Convert phoneme to waveform
  - - - Smoothing
        
        Removing discontinuities at the boundaries of concatenated segments.
      - Filtering
        
        Adjusting the audio signal to improve clarity and naturalness
      - Normalization
        
        Adjusting the volume level.
- - - - Data augmentation
        
        Techniques like time stretching, pitch shifting, and adding background noise artificially enrich the training data, improving model robustness to variations in real-world speech
      - Text normalization
  - - - Poor Prosody
        
        Unnatural or monotonous intonation, rhythm, and stress patterns
      - Mispronunciation
        
        Homograph Errors
        
        Incorrect pronunciation of words that have the same spelling but different pronunciations based on context
        
        Homo-phoneme Errors
        
        Incorrect pronunciation of words that have the same pronunciation but different spelling
      - Distortion and artifact
    - - Deep learning architectures
        
        LSTM
        
        transformers
      - Multi-speaker training
        
        Training on data from multiple speakers enables the model to generate diverse voices and adapt to different speaking styles.
      - Prosody modeling
        
        Techniques like phrase boundaries and intonation control regulate the rhythm, emphasis, and pitch of the synthesized speech, improving naturalness and clarity.
      - End-to-end learning
        
        This approach integrates text processing and speech synthesis in a single model, simplifying the pipeline and potentially improving both accuracy and naturalness
  - - - Overfitting to a specific speaker
      - Misadapted pronunciation or prosody for specific languages
    - - Voice modification
      - Speaker adaptation
        
        voice filter with few shot taring model
        
        Multi Speaker Training
      - speech enhancements
        
        speech denoising
        
        dereverberation
        
        speech super-resolution
- - - - According to Weerasighe et al. [9], their synthesis system
        named "Festival-si
      - Nanayakkarara el al. [10] proposes a unit selection mechanism
        using MarryTTS framework to
      - Wickramasinghe, Kumara, and Dias [12]
        highlight several fundamental challenges
        
        Speech and text corpora are scarcely available
        
        Code-switching inside text
        
        Confusion between numbers, fractions, and dates within
        the text
        
        Handling special characters within the text is difficult
        
        Abbreviations within the text
- - - - Acoustic modeling
      - Vocoding