FireRedASR

Why we included this project

FireRedASR is an open-source, industrial-grade ASR family with two ready-to-run model variants: an efficient encoder-decoder model and a larger LLM-based one, so you can trade latency and memory against accuracy depending on where you deploy. It transcribes Mandarin, Chinese dialects, and English, with error rates on public Mandarin benchmarks among the best reported, which matters when transcription feeds search, live captioning, or downstream language models. The singing lyrics recognition is a niche extra that few mainstream toolkits attempt. Setup is plain: download weights from Hugging Face, convert audio to 16 kHz PCM, then use the CLI or Python API for files or batches. Just note the AED variant handles about 60 seconds of audio per call and the LLM variant about 30, so split longer recordings first.