TTS Concepts

Pipeline and Architecture

• TTS Pipeline (End-to-End) — How text flows through G2P, acoustic model, and vocoder.
• Formant vs Parametric vs Neural TTS — The three generations of speech synthesis.
• Autoregressive vs Non-Autoregressive TTS — Two neural architectures with different speed and quality tradeoffs.
• Diffusion Models for TTS — The latest architecture paradigm in speech synthesis.
• Attention and Alignment — How TTS models learn to align text to audio.

Text and Pronunciation

• G2P (Grapheme-to-Phoneme) — Converting written text into phonetic representations.
• Text Tokenization — How raw text is split into tokens for TTS models.
• SSML (Speech Synthesis Markup Language) — XML markup for controlling TTS output.

Voice and Quality

• Prosody — The rhythm, stress, and intonation of synthetic speech.
• Voice Cloning and Speaker Embedding — Replicating a voice from audio samples.
• Zero-Shot vs Few-Shot vs Fine-Tuned TTS — Three approaches to voice adaptation.
• Emotion and Expressiveness — How TTS handles emotional delivery.

Signal Processing

• Mel-Spectrograms — The frequency representation used by neural TTS models.
• Vocoder — Converting acoustic features into audio waveforms.

Generation and Performance

• Duration Prediction — How TTS determines phoneme timing and pacing.
• Chunking — Splitting long text into segments for TTS generation.
• Streaming TTS — Generating speech incrementally for real-time applications.
• Real-Time Factor (RTF) — The standard metric for TTS generation speed.
• MOS (Mean Opinion Score) — How TTS voice quality is measured.