FireRedASR2 AED-L · OpenASR
FireRedTeam's Mandarin-first bilingual ASR — attention encoder-decoder tuned for state-of-the-art Chinese and dialect accuracy
Native speech-to-text in the
OpenASR runtime —
engineered for peak performance on CPU & GPU,
no Python at inference time.
✨ Highlights
- 🥇 State-of-the-art Mandarin accuracy — 3.05% average CER across four public Mandarin benchmarks, with 0.57% CER on aishell1 (arXiv:2603.10420 Table 2)
- 🗣️ Strong dialect coverage — 11.67% average CER across 19 public Chinese dialect/accent benchmarks, beating Doubao-ASR (15.39%) and matching Qwen3-ASR (11.85%) on the paper's own comparison table
- 🇨🇳🇬🇧 Bilingual Mandarin + English — one 1.1B checkpoint decodes both languages plus code-switching, no language flag needed
- 🎤 Singing-lyrics robustness — the paper reports 1.17% CER on a singing-lyrics benchmark, alongside the speech-recognition numbers above
- 🏗️ Conformer encoder + Transformer decoder — an attention-based encoder-decoder, architecturally identical to FireRedASR-AED-L, retrained on a larger vocabulary
- 🦀 Native in OpenASR —
.oasr packs run with no Python at inference, engineered for peak performance on CPU & GPU
🚀 Quickstart
1# 1. Install the OpenASR CLI · https://openasr.org
2# 2. Pull a build (pick a quant — see the table below)
3openasr pull firered-aed-l-v2:q4
4
5# 3. Transcribe
6openasr transcribe audio.wav --model firered-aed-l-v2
All builds for this model:
1openasr pull firered-aed-l-v2:fp16
2openasr pull firered-aed-l-v2:q8
3openasr pull firered-aed-l-v2:q4
📦 Available builds
| Quant | File (.oasr) | Size | RAM peak | RTF · M1 CPU | RTF · M1 GPU | JFK ΔWER vs fp16 |
|---|
| fp16 | firered-aed-l-v2-fp16.oasr | 2.35 GB | 4.74 GB | 0.40× | 0.79× | 0.0% |
| q8_0 | firered-aed-l-v2-q8_0.oasr | 1.28 GB | 4.35 GB | 0.36× | 0.35× | 0.0% |
| q4_k | firered-aed-l-v2-q4_k.oasr | 1.06 GB | 1.27 GB | n/a | n/a | n/a |
RTF = real-time factor on the fixed 11s JFK clip (lower is faster); RAM peak measured per pack
in an isolated subprocess. JFK ΔWER compares each quantized build's JFK transcript to this model's
fp16 JFK transcript, so it measures quantization drift rather than absolute recognition accuracy.
q4_k is the recommended default — near-reference quality at a fraction of the
footprint.
🧠 About FireRedASR2 AED-L
FireRedASR2-AED is the attention-based encoder-decoder member of FireRedASR2, the successor
to FireRedTeam's open-source industrial-grade FireRedASR speech-recognition family, released
as part of the FireRedASR2S all-in-one ASR system. It pairs a Conformer encoder with a
Transformer decoder at 1.1B parameters -- architecturally byte-identical to the original
FireRedASR-AED-L -- retrained on a larger token vocabulary. The FireRedASR2S technical report
(arXiv:2603.10420, Table 2) reports 3.05% average Character Error Rate across four public
Mandarin benchmarks (aishell1 at 0.57%), 11.67% average CER across 19 public Chinese
dialect/accent benchmarks, and 1.17% CER on a singing-lyrics benchmark, outperforming
Doubao-ASR and matching Qwen3-ASR on the paper's own comparison table. It is bilingual (Mandarin
Chinese and English, including code-switching). This OpenASR repo repackages the weights as
.oasr packs that run natively in the OpenASR runtime -- no Python at inference time, all
decoding local. The q4_k build is the recommended default for everyday use; q8_0 trades
some size for closer fp16 fidelity and fp16 is for maximum fidelity or verification. Note:
FireRedTeam also publishes a larger FireRedASR2-LLM (Encoder-Adapter-LLM, 2.89%/11.55% avg
CER) variant that this repo does not distribute -- the numbers above are for this AED checkpoint
only, and should not be conflated with the original (v1) FireRedASR-AED-L's 3.18%/0.55% figures,
which come from a different training run and benchmark protocol.
⚙️ How these packs were made
Converted from
FireRedTeam/FireRedASR2-AED with the OpenASR importer:
1openasr model-pack import firered-aed <src> <out>.oasr \
2 --package-id firered-aed-l-v2 --quantization {fp16,q8-0,q4-k}
The .oasr container is GGUF-backed; packs use zero-copy mmap weight binding and graph
buffer reuse to keep peak memory low.
⚖️ License
These packs
inherit the upstream model's license: Apache-2.0
(
source). OpenASR packaging retains the upstream copyright and
NOTICE; the only modifications are format conversion and quantization.
🙏 Acknowledgements
This pack is a redistribution of
FireRedASR2-AED, created and released by
FireRedTeam
(
FireRedTeam/FireRedASR2-AED,
FireRedTeam/FireRedASR2S). All credit for the
architecture, training, and weights belongs to FireRedTeam; the license is inherited from and
identical to the upstream model (
Apache-2.0, as declared on the upstream model card). Thank
you to FireRedTeam for releasing their work openly. OpenASR only performs format conversion,
quantization, runtime verification, and local-inference adaptation.
🔗 Links
