NeuTTS-2E: An Open-Source On-Device TTS Model with Emotion Control
Neuphonic has released the alpha version of NeuTTS-2E, a 125-million-parameter on-device text-to-speech (TTS) model. This open-source model allows users to explicitly control seven different emotions while maintaining speaker consistency. This release enables developers to deploy highly expressive, private, and local speech synthesis on consumer hardware without relying on cloud APIs. By allowing direct emotional control, it solves the issue where TTS models only infer emotion from the semantic meaning of the input text. NeuTTS-2E supports four built-in voices, runs entirely locally, and is released under the NeuTTS Open License. To achieve this, the developers had to overcome challenges such as limited emotional speech training data and the difficulty of disentangling spoken emotion from text semantics.
## BACKGROUND
Text-to-speech (TTS) synthesis is the artificial production of human speech from written text. A major challenge in emotional speech synthesis is disentangling speaker identity and text semantics from emotional expression, which often causes voices to sound unnatural or lose their identity when changing tones. On-device TTS models are increasingly popular because they ensure user privacy and run without internet latency.