Google’s Tacotron 2 Revolutionizes AI Speech Synthesis
Google’s latest innovation, Tacotron 2, marks a considerable leap forward in artificial speech generation. This system produces remarkably natural-sounding speech from text, unlike previous methods, requiring significantly less linguistic data for training. Tacotron 2 cleverly combines the strengths of two earlier Google projects: WaveNet and the original Tacotron. WaveNet excelled at generating high-quality audio but demanded extensive linguistic data. Tacotron focused on high-level speech features like intonation and prosody but fell short in producing a final, polished audio product.
Tacotron 2 leverages both approaches, using text and its narration to infer linguistic rules and create remarkably natural-sounding audio. While the audio quality is impressive in most cases, challenges remain. The system occasionally struggles with the pronunciation of certain words, particularly those of non-English origin, and lacks control over the overall tone of the speech. The reduced training data needs, however, are a major breakthrough, enabling the creation of superior systems and easier integration of new approaches without extensive manual rule adjustments or the need for new rulesets for various languages and speaking styles. The research has been submitted to the IEEE International Conference on Acoustics, Speech and Signal Processing and is available on arXiv.
