FastSpeech Series by Microsoft

FastSpeech is a family of non-autoregressive text-to-speech models developed by Microsoft. The second iteration, FastSpeech 2, introduced enhancements in voice quality, training efficiency, and controllability.

• Utilizes a feed-forward Transformer architecture for efficient parallel mel-spectrogram generation

• Incorporates pitch, energy, and duration predictors to improve prosody and expressiveness

• Achieves faster training and inference compared to autoregressive models

• Supports multi-speaker and multilingual synthesis with appropriate training data

• Open-source implementations available in PyTorch and TensorFlow

CLICK HERE TO DISCOVER FASTSPEECH SERIES