Audio samples from "Controllable neural text-to-speech synthesis using intuitive prosodic features"
Authors: Tuomo Raitio, Ramya Rasipuram, Dan Castellani
Abstract: 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).
The following samples demonstrate the prosody control capability by adjusting bias of each prosodic dimension.
Baseline: High-quality baseline model trained with 36 hours of normal speech data
[No control]
-
-
Prosody 36h: Prosody control model trained with 36 hours of normal speech data
Feature / Bias
-1.0
-0.7
-0.5
-0.3
0.0
0.3
0.5
0.7
1.0
Pitch
Pitch range
Duration
Energy
Spectral tilt
Prosody 52h: Prosody control model trained with 36 hours of normal speech data + 16 hours of expressive speech data
Feature / Bias
-1.0
-0.7
-0.5
-0.3
0.0
0.3
0.5
0.7
1.0
Pitch
Pitch range
Duration
Energy
Spectral tilt
The following samples demonstrate the capability to create meaningfully different versions of the same utterance.