Controllable neural text-to-speech synthesis using intuitive prosodic features
This work addresses the need for more expressive and controllable speech synthesis for applications like virtual assistants and audiobooks, representing an incremental improvement over existing methods.
The paper tackled the problem of limited prosodic variation in neural text-to-speech synthesis by training a sequence-to-sequence model conditioned on acoustic features to learn a controllable latent prosody space, resulting in a model that maintains a mean opinion score of 4.23 compared to a baseline of 4.26 while enabling diverse speaking styles.
Modern neural text-to-speech (TTS) synthesis can generate speech that is indistinguishable from natural speech. However, the prosody of generated utterances often represents the average prosodic style of the database instead of having wide prosodic variation. Moreover, the generated prosody is solely defined by the input text, which does not allow for different styles for the same sentence. In this work, we train a sequence-to-sequence neural network conditioned on acoustic speech features to learn a latent prosody space with intuitive and meaningful dimensions. Experiments show that a model conditioned on sentence-wise pitch, pitch range, phone duration, energy, and spectral tilt can effectively control each prosodic dimension and generate a wide variety of speaking styles, while maintaining similar mean opinion score (4.23) to our Tacotron baseline (4.26).