This article provides a practical comparison between two major open-source text-to-speech models from Alibaba: CosyVoice3 and IndexTTS2. The test involved cloning voiceovers from characters in the Arknights game and comparing them with the original human voiceovers. Results show that IndexTTS2 performs better in terms of speech naturalness, coming close to the original voice effect. Meanwhile, CosyVoice3 has significant advantages in inference speed and resource consumption, generating an audio clip in just about 10 seconds—much faster than IndexTTS2’s minute and a half. The article notes that CosyVoice3 supports direct natural language control and phoneme methods, optimizing synthesized text through auxiliary small models without significantly compromising quality. For readers interested in AI speech synthesis technology, this comparison offers practical guidance for model selection in different scenarios.
