We introduce QinYu, a family of high-fidelity text-to-speech systems capable of generating speech 32kHz studio-quality speech with exceptional naturalness. For the audiobook scenario, we have achieved fine-grained control of emotions through text instructions described in natural language, which significantly enhances emotional expressiveness. Meanwhile, we have also researched text-to-timbre technology: by describing the gender, age, and personality of a desired timbre, we can generate the corresponding voice timbre, thus solving the problems of limited timbre options and difficulties in timbre matching. For the podcast dialogue scenario, we have implemented the ability for spontaneous colloquial expression (with automatic addition of pauses, hesitations, and moments of thinking) as well as enhanced paralinguistic expression, resulting in a more realistic and human-like effect.The goal of systems is to close the gap between synthetic and human speech across diverse TTS applications.
