FireRedASR2S

Why we included this project

FireRedASR2S suits teams that need to transcribe Chinese audio, especially when the speech mixes dialects, code-switching, or singing. Rather than a single recognizer, it ships four modules: ASR, voice activity detection, language identification, and punctuation restoration, so a full transcription chain can be assembled from one codebase. The ASR core covers Mandarin plus 20+ Chinese dialects and accents, English, and code-switched speech, which is where generic Mandarin models tend to fall short in call centers, media archives, and meeting recordings. The AED variant adds word-level timestamps and confidence scores, handy for alignment and quality filtering downstream. vLLM and TensorRT-LLM inference paths, along with streaming VAD, make it deployable in latency-sensitive services. If your audio is mostly Chinese with dialect variety, it is worth benchmarking against the usual ASR options.