Standard TTS uses concatenative synthesis. It stitches pre-recorded phonetic fragments together. Sounds robotic because the intonation and connection between fragments is rule-based.
Neural TTS uses a deep learning model (like Tacotron 2, WaveNet) trained on hours of human speech. It doesn't stitch; it generates the raw audio waveform from the text, predicting things like prosody and emphasis in a more holistic way.
Key technical differences:
* **Input/Output:** Concatenative = text -> select units -> concatenate. Neural = text -> acoustic features -> waveform (via neural vocoder).
* **Prosody:** Neural models capture subtle variations (pitch, timing, stress) from training data, making it sound more natural and less monotone.
* **Data:** Requires massive, clean, labeled speech datasets for training.
Think of it as the difference between playing a MIDI file (concatenative) versus a live musician interpreting the score (neural). The latter has nuance and flow that's hard to codify with rules.
-dk
Trust but verify, then don't trust.