告別AI機器音！Index TTS2 讓你的聲音充滿喜怒哀樂，真正實現聲音自由！
Large-scale text-to-speech (TTS) models are typically categorized into autoregressive and non-autoregressive systems. Although autoregressive systems exhibit certain advantages in speech naturalness, their token-by-token generation mechanism makes it difficult to precisely control the duration of the synthesized speech. This becomes a significant limitation in applications such as video dubbing, where strict audio-visual synchronization is required. This paper introduces IndexTTS2, which proposes a novel, general, and autoregressive-model-friendly method for speech duration control. The method supports two generation modes: one allows explicit specification of the number of generated tokens, thereby enabling precise control over speech duration; the other does not require manual token count input, letting the model freely generate speech in an autoregressive manner while faithfully reproducing prosodic characteristics from the input prompt. Furthermore, IndexTTS2 achieves disentanglement between emotional expression and speaker identity, enabling independent control of timbre and emotion. In the zero-shot setting, the model is capable of perfectly reproducing the emotional characteristics inherent in the input prompt. Additionally, users may provide a separate emotion prompt (which can originate from a different speaker than the timbre prompt), thereby enabling the model to accurately reconstruct the target timbre while conveying the specified emotional tone. In order to enhance the clarity of speech during strong emotional expressions, we incorporate GPT latent representations to improve the stability of the generated speech. Meanwhile, to lower the barrier for emotion control, we design a soft instruction mechanism based on textual descriptions by fine-tuning Qwen3. This facilitates the effective guidance of speech generation with the desired emotional tendencies through natural language input. Finally, experimental results on multiple datasets demonstrate that IndexTTS2 outperforms existing state-of-the-art zero-shot TTS models in terms of word error rate, speaker similarity, and emotional fidelity. To promote further research and facilitate practical adoption, we will release both the model weights and inference code, enabling the community to reproduce and build upon our work.
留言
張貼留言
歡迎留言