Shijia Liao is founder and chief scientist of Fish Audio. He works on speech-generation models and the training and inference systems that make expressive voice output controllable and practical to serve.[1][3][5]
Start with How we made our text-to-speech API free →
Education
Links
Fish Audio author page LinkedIn
X · LengyueMatrix Founder and Chief Scientist · Fish Audio
Career
Contributions
Describe the performance as well as the words
Liao is first-listed co-author of the Fish Audio S2 report. The system combines multi-speaker, multi-turn speech generation with natural-language instructions for delivery. Its training recipe includes staged captioning, voice-quality assessment and reward modeling. The team releases weights, fine-tuning code and an inference engine; benchmark results describe the reported setup rather than every deployment.[3]
Optimize the small operations inside the decode loop
His inference-engineering article explains why autoregressive speech generation can leave GPU compute idle: small matrix operations and repeated memory access dominate each step. The team uses specialized kernels, reduced-precision computation and continuous batching to improve the workload. The account connects those implementation choices with the cost of serving speech, while reporting measurements for specific hardware and batch conditions.[5]
Extend directed speech into longer performances
He also co-authored the Drama 3 technical report. It studies directed speech generation with a mixture-of-experts model, instruction control and training and evaluation methods for expressive performance. This is a larger team effort, with the report’s author list and experimental conditions defining the contribution.[4]
Tools, writing & talks
Writing, research & talks
Start here
How we made our text-to-speech API freefish.audioAn engineering account of decode kernels, batching and inference cost for speech generation.
Start with the decode bottleneck and why batching helps before the kernel details. Familiarity with GPU inference is useful; the measurements and temporary API pricing belong to the article’s stated conditions.
Author[5]- Fish Audio S2 technical reportarxiv.org
Training and inference methods for controllable, multi-speaker speech generation.
First-listed co-author[3] - Fish Audio Drama 3 technical reportarxiv.org
A team report on directed expressive speech generation and evaluation.
Co-author[4]