Huggingface-Aishell get wrong prediction

[Ht-zhang-xianyu]

Describe the bug

Use the pretrained from huggingface case in Aishell dataset, but the prediction seems strange
The script I use is as follows:

`
from speechbrain.inference.ASR import EncoderDecoderASR

asr_model = EncoderDecoderASR.from_hparams(source="speechbrain/asr-transformer-aishell", savedir="pretrained_models/asr-transformer-aishell")
asr_model.transcribe_file("speechbrain/asr-transformer-aishell/example_mandarin2.flac")`

but return the predict token is :
"一 日 一一 一一 一一 六 克一 件 第 一"

Expected behaviour

expect the correct asr script of wav

To Reproduce

No response

Environment Details

No response

Relevant Log Output

No response

Additional Context

No response

[LudwigFancy]

i meet the same question, then i use this pretrained model to test another examples which is created by myself, but more errors happened. i don't understand whether the data need a preprocess when we test?(I'm a novice at this)

[asumagic]

Seems like develop is even more broken due to the mutihead_attn rename, resulting in some layer weights not being loaded correctly:

loading from pretrained_models/asr-transformer-aishell/asr.ckpt, the object could not use the parameters loaded with the key: 1.decoder.layers.5.mutihead_attn.att.out_proj.bias

#2529 solves that.

That being said, the current issue above is separate.