Neural TTS Emphasis via Acoustic Embeddings

The patent embeds multiple levels of neural network models within each other - a phoneme-level emphasis detection model is nested within a word-level model, which is nested within a sentence-level prosody generation model. This nested architecture allows the system to capture emphasis at different granularities while maintaining computational efficiency through shared latent representations.

The patent transforms the one-dimensional text input into multi-dimensional acoustic embeddings that encode both linguistic content and prosodic features. By projecting text into a higher-dimensional acoustic space, the system can simultaneously represent phoneme identities, emphasis patterns, and prosodic contours, enabling natural-sounding synthesis without requiring separate processing modules.

The patent performs emphasis detection and acoustic embedding generation in advance during the text processing stage, before actual speech synthesis. By pre-computing the emphasis patterns and encoding them into acoustic embeddings, the system avoids real-time computational overhead during speech generation, maintaining both accuracy and efficiency.

The patent uses pre-trained neural network models that have learned emphasis patterns from large corpora of natural speech. These models copy proven effective prosodic patterns from training data and apply them to new text inputs, avoiding the need to compute emphasis from scratch and significantly reducing processing time while maintaining high accuracy.