Spark-TTS

Spark-TTS rebuilds text to speech directly on top of a language model: a Qwen2.5-based 0.5B model predicts audio codes that the single-stream BiCodec codec decodes straight to waveform, removing the separate acoustic-model and vocoder stages most systems chain together. The approach supports zero-shot voice cloning from a short reference clip plus controllable generation, letting users set gender, pitch, and speaking rate to create entirely virtual voices, and it handles Chinese and English including cross-lingual and code-switched text. The SparkAudio team published the method as arXiv 2503.01710 in March 2025 and released weights on Hugging Face; the repository documents Python 3.12, PyTorch 2.5, and single-GPU inference, and adds an NVIDIA Triton and TensorRT-LLM runtime benchmarked on a single L20 at 876 ms average latency and a real-time factor of 0.14 at concurrency one. Code is Apache-2.0 with a disclaimer restricting misuse for impersonation or fraud. Since its March 2025 release it has passed 11,000 GitHub stars and become a common base for voice cloning research and self-hosted speech pipelines.