Problems encountered while fine-tuning my own model

[Nerva05251228]

As shown in the title, after using the following code to fine-tune the model, I found that the fine-tuned model recognized any audio as [''], including when I used the original language as input, but it was still blank after recognition. I don't know what happened?
The code is as follows:

import nemo
import nemo.collections.asr as nemo_asr

train_manifest = "/root/autodl-tmp/nemo/json/Yue/train_1st.json"
test_manifest = "/root/autodl-tmp/nemo/json/Yue/test_1st.json"

try:
    from ruamel.yaml import YAML
except ModuleNotFoundError:
    from ruamel.yaml import YAML
config_path = "/root/autodl-tmp/nemo/config/quartznet_15x5_zh.yaml"

yaml = YAML(typ="safe")
with open(config_path) as f:
    params = yaml.load(f)
print(params)

params['model']['train_ds']['manifest_filepath'] = train_manifest
params['model']['validation_ds']['manifest_filepath'] = test_manifest

quartznet.setup_training_data(train_data_config=params['model']['train_ds'])
quartznet.setup_validation_data(val_data_config=params['model']['validation_ds'])
import pytorch_lightning as pl
trainer = pl. Trainer(gpus=1, max_epochs = 200)
trainer.fit(quartznet)

quartznet.save_to("1st_asr_model_Yue.nemo")

try_model_1 = nemo_asr.models.EncDecCTCModel.restore_from("1st_asr_model_Yue.nemo")

asr_result_1 = try_model_1.transcribe(paths2audio_files=["/root/autodl-tmp/nemo/source/test/Yue/voice3.wav"])
print(asr_result_1)

Here is the output:

Transcribing: 100%  1/1 [00:00<00:00, 10.10it/s]
['']

By the way, I used a Mandarin model as the base and used over twenty hours of Cantonese audio, with a training epochs of 200

[Nerva05251228]

Over fitting or under fitting? Should I reduce epochs or reduce batch_ size?