Tacotron 2 Improvements
What you'll learn: How Tacotron 2 enhances the original architecture with location-sensitive attention and a WaveNet vocoder to produce near-human-quality synthesized speech.
The Problem with Tacotron 1
While Tacotron pioneered end-to-end text-to-speech, it had limitations: the attention mechanism sometimes "wandered" or repeated words, and the Griffin-Lim vocoder (converting mel spectrograms to audio) produced muffled output.
Location-Sensitive Attention
Standard attention in Tacotron treats each decoding step independently—it forgets where it attended previously. This causes the model to lose track during long sentences.
Location-sensitive attention adds positional awareness by feeding information about previous attention alignments back into the current step. Think of it like reading with your finger: you remember where you just were, making it easier to move smoothly to the next word without skipping or repeating.
The mechanism concatenates:
- Previous attention weights (where you just looked)
- Convolutional features extracted from the attention history
This guides the decoder to progress naturally left-to-right through the input text, dramatically reducing attention errors.
WaveNet Vocoder
Instead of Griffin-Lim, Tacotron 2 uses a WaveNet vocoder to convert mel spectrograms into raw audio waveforms. WaveNet is a deep convolutional network that models audio one sample at a time, capturing subtle acoustic details that simpler vocoders miss.
The result? Speech that sounds remarkably natural—almost indistinguishable from human recordings in quality tests.
Why It Matters
These two improvements work in tandem: location-sensitive attention ensures the model generates correct, coherent mel spectrograms, while WaveNet transforms them into crisp, natural-sounding audio. Together, they elevated TTS from "robotic" to "nearly human."
Key Takeaway: Tacotron 2 achieves state-of-the-art speech quality by adding location-sensitive attention for robust alignment and replacing simple vocoders with WaveNet for high-fidelity audio generation.