FireRedASR

What it is

FireRedASR is a family of open-source industrial-grade automatic speech recognition models supporting Mandarin, Chinese dialects and English, including two variant architectures: FireRedASR-LLM (Encoder-Adapter-LLM) for end-to-end speech interaction and FireRedASR-AED (Attention-based Encoder-Decoder) for efficient speech representation in LLM-based systems. The project provides model weights, inference code, and evaluation data.

How it works

FireRedASR consists of two architectures:

• FireRedASR-LLM: Encoder-Adapter-LLM framework leveraging large language model capabilities to achieve end-to-end performance.
• FireRedASR-AED: AED architecture balancing performance and computational efficiency for use as a module within LLM-based systems. Evaluation reports CER% for Chinese and WER% for English across several benchmarks. The README lists comparisons against multiple baselines on public Mandarin and English ASR benchmarks.

Getting started

Install and run as described:

$ git clone https://github.com/FireRedTeam/FireRedASR.git
$ conda create --name fireredasr python=3.10
$ conda activate fireredasr
$ pip install -r requirements.txt

Set PATH and PYTHONPATH:

$ export PATH=$PWD/fireredasr/:$PWD/firreredasr/utils/:$PATH
$ export PYTHONPATH=$PWD/:$PYTHONPATH

Convert audio to 16kHz 16-bit PCM:

ffmpeg -i input_audio -ar 16000 -ac 1 -acodec pcm_s16le -f wav output.wav

Quick Start:

$ cd examples
$ bash inference_fireredasr_aed.sh
$ bash inference_fireredasr_llm.sh

Command-line usage:

$ speech2text.py --help
$ speech2text.py --wav_path examples/wav/BAC009S0764W0121.wav --asr_type "aed" --model_dir pretrained_models/FireRedASR-AED-L
$ speech2text.py --wav_path examples/wav/BAC009S0764W0121.wav --asr_type "llm" --model_dir pretrained_models/FireRedASR-LLM-L

Python usage:

from fireredasr.models.fireredasr import FireRedAsr
batch_uttid = ["BAC009S0764W0121"]
batch_wav_path = ["examples/wav/BAC009S0764W0121.wav"]

# FireRedASR-AED
model = FireRedAsr.from_pretrained("aed", "pretrained_models/FireRedASR-AED-L")
results = model.transcribe(
    batch_uttid,
    batch_wav_path,
    {
        "use_gpu": 1,
        "beam_size": 3,
        "nbest": 1,
        "decode_max_len": 0,
        "softmax_smoothing": 1.25,
        "aed_length_penalty": 0.6,
        "eos_penalty": 1.0
    }
)
print(results)

# FireRedASR-LLM
model = FireRedAsr.from_pretrained("llm", "pretrained_models/FireRedASR-LLM-L")
results = model.transcribe(
    batch_uttid,
    batch_wav_path,
    {
        "use_gpu": 1,
        "beam_size": 3,
        "decode_max_len": 0,
        "decode_min_len": 0,
        "repetition_penalty": 3.0,
        "llm_length_penalty": 1.0,
        "temperature": 1.0
    }
)
print(results)

Recent releases

• 2026.02.25: FireRedASR2-LLM model weights release and related updates.
• 2026.02.12: FireRedASR2S release including AED, VAD, LID, and Punc with model weights and inference code.
• 2025.02.17: FireRedASR-LLM-L model weights release.
• 2025.01.24: Technical report, blog, and AED-L model weights release.

Traction

stars_7d: not provided stars_1d: not provided

Behind the repo

FireRedASR2S is linked as a follow-up project with additional modules such as VAD, LID, and Punc, indicating an ongoing development track within the same organization.

Caveats

License: Apache-2.0 Created: 2025-01-24 Last push: 2026-02-25 Usage notes mention input length limitations: AED supports up to 60s; LLM up to 30s. Open issues are listed as 52 on the repo.