Modelpublic

inoryQwQ/sherpa-onnx-fire-red-asr-ax650

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes8downloads

Model Card

sherpa-onnx-fire-red-asr-ax650

Fire Red ASR speech recognition model converted to AX650 AXMODEL for on-device inference via sherpa-onnx.

Model Details

—Original Model: FireRedTeam/FireRedASR-LLM-L
—Architecture: Encoder-Decoder with BPE tokenizer
—Task: Automatic Speech Recognition (ASR)
—Sample rate: 16000 Hz
—Target Chip: AX650 (NPU3)
—Quantization: U16

Files

File	Size	Description
`encoder.axmodel`	812 MB	Encoder AXMODEL
`decoder_loop.axmodel`	397 MB	Decoder loop AXMODEL
`train_bpe1000.model`	246 KB	BPE tokenizer model (1000 tokens)
`dict.txt`	70 KB	Dictionary / vocabulary file
`cmvn.ark`	1.3 KB	CMVN normalization parameters
`pe.npy`	24 MB	Position encoding weights

Usage with sherpa-onnx

bash

./sherpa-onnx-offline \
  --fire-red-asr-encoder=encoder.axmodel \
  --fire-red-asr-decoder=decoder_loop.axmodel \
  --tokens=dict.txt \
  --bpe-model=train_bpe1000.model \
  --provider=axera \
  audio.wav

Performance

Metric	Value
AXMODEL size (total)	1.21 GB
RTF	TBD

Conversion Details

—Pulsar2 Version: 6.0
—Calibration: MinMax with 10 samples
—Input Shape: encoder [1, T, 80], decoder [1, T, 512]

License

Apache 2.0