FireRedASR

FireRedASR comes from the FireRed team at Xiaohongshu and packages two industrial-grade speech recognition architectures for Mandarin, Chinese dialects, and English, including recognition of sung lyrics. FireRedASR-AED is a 1.1B parameter attention-based encoder-decoder built for latency-sensitive production use, while FireRedASR-LLM couples an audio encoder to a large language model through an Encoder-Adapter-LLM design at 8.3B parameters for maximum accuracy. On public Mandarin benchmarks the LLM variant averages a 3.05 percent character error rate across four test sets with 0.76 percent on AISHELL-1, and the AED variant averages 3.18 percent, results the authors describe as a new state of the art (arXiv 2501.14350). Setup means cloning the repository, creating a Python 3.10 environment, installing pip dependencies, and downloading weights from Hugging Face into a pretrained_models folder; inference realistically wants an NVIDIA GPU, especially for the 8.3B model. Code and released models carry the Apache-2.0 license, so commercial deployment is permitted.