Linguistic Features and Text Processing
You'll learn how to transform raw input text into meaningful linguistic features that guide neural speech synthesis systems.
Why Text Processing Matters for Speech Synthesis
Before a TTS system can speak, it must understand the text. Raw characters like "read" could mean /riːd/ (present tense) or /rɛd/ (past tense). Your neural vocoder needs explicit guidance about pronunciation, rhythm, and emphasis—that's where linguistic feature extraction comes in.
Core Components
Phoneme Extraction
Convert written text into phonemes—the atomic sound units. The word "cat" becomes /k/ /æ/ /t/. This involves:
- Grapheme-to-phoneme (G2P) conversion: mapping spelling to sounds
- Handling homophones and context-dependent pronunciations
- Using pronunciation dictionaries or learned G2P models
Prosody Markers
Prosody controls the "melody" of speech—pitch, duration, and stress patterns. Extract:
- Syllable stress: which syllables are emphasized ("REcord" vs "reCORD")
- Phrase boundaries: where to pause for commas, periods
- Question intonation: rising pitch at sentence end
Additional Linguistic Features
- Part-of-speech tags: nouns vs verbs affect pronunciation
- Punctuation encoding: commas signal brief pauses
- Word boundaries: help the model segment properly
Think of this stage as translating English into "speech notation"—a recipe the synthesis model can follow precisely.
From Text to Neural Input
These extracted features become input vectors fed into your acoustic model (the component that predicts mel-spectrograms). Instead of random character embeddings, you're giving the network linguistic knowledge.
Key Takeaway: Linguistic feature extraction transforms ambiguous text into explicit phonetic, prosodic, and structural representations that enable accurate, natural-sounding speech synthesis.