FireRedASR

Open-source industrial-grade Mandarin speech recognition from Xiaohongshu's FireRed team, in two flavors: FireRedASR-AED-L (1.1B, attention-based encoder-decoder, efficiency-focused) and FireRedASR-LLM-L (8.3B, Encoder-Adapter-LLM framework), achieving SOTA on public Mandarin benchmarks at release, with support for Chinese dialects and English. Succeeded by FireRedASR2 (February 2026), an all-in-one ASR/VAD/LID/punctuation system. Apache-2.0 licensed.