publish firered-aed-l-v2 OpenASR packs

dc176bb verified 1 day ago

6.54 kB

	---
	license: apache-2.0
	base_model: FireRedTeam/FireRedASR2-AED
	pipeline_tag: automatic-speech-recognition
	library_name: openasr
	tags:
	- automatic-speech-recognition
	- speech-to-text
	- openasr
	- oasr
	- firered-aed-l-v2
	---

	<div align="center">

	# FireRedASR2 AED-L · OpenASR

	FireRedTeam's Mandarin-first bilingual ASR — attention encoder-decoder tuned for state-of-the-art Chinese and dialect accuracy

	[![License](https://img.shields.io/badge/license-Apache--2.0-2563eb.svg)](https://huggingface.co/FireRedTeam/FireRedASR2-AED/blob/main/README.md)
	[![Format](https://img.shields.io/badge/format-.oasr-7c3aed.svg)](https://github.com/QuintinShaw/openasr)
	[![Runtime](https://img.shields.io/badge/runtime-OpenASR-111827.svg)](https://openasr.org)
	[![Base model](https://img.shields.io/badge/base-FireRedASR2--AED-f59e0b.svg)](https://huggingface.co/FireRedTeam/FireRedASR2-AED)

	Native speech-to-text in the [OpenASR](https://github.com/QuintinShaw/openasr) runtime —
	engineered for peak performance on CPU & GPU, no Python at inference time.

	</div>

	---

	## ✨ 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

	```bash
	# 1. Install the OpenASR CLI · https://openasr.org
	# 2. Pull a build (pick a quant — see the table below)
	openasr pull firered-aed-l-v2:q4

	# 3. Transcribe
	openasr transcribe audio.wav --model firered-aed-l-v2
	```

	All builds for this model:

	```bash
	openasr pull firered-aed-l-v2:fp16
	openasr pull firered-aed-l-v2:q8
	openasr 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` \| 709 MB \| 4.10 GB \| 0.38× \| 0.31× \| 0.0% \|

	<sub>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.</sub>

	## 🧠 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](https://huggingface.co/FireRedTeam/FireRedASR2-AED) with the OpenASR importer:

	```bash
	openasr model-pack import firered-aed <src> <out>.oasr \
	--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](https://huggingface.co/FireRedTeam/FireRedASR2-AED/blob/main/README.md)). 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](https://huggingface.co/FireRedTeam/FireRedASR2-AED),
	[FireRedTeam/FireRedASR2S](https://github.com/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

	- 🦀 OpenASR — <https://github.com/QuintinShaw/openasr>
	- 🌐 Website — <https://openasr.org>
	- 🤗 Upstream model — [FireRedTeam/FireRedASR2-AED](https://huggingface.co/FireRedTeam/FireRedASR2-AED)