如何使用ModelScope魔搭训练工业级语音识别模型

from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks

inference_16k_pipline = pipeline(
task=Tasks.auto_speech_recognition,
model='damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch')

rec_result = inference_16k_pipline(audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav')
print(rec_result)

rec_result = inference_16k_pipline(audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.pcm', audio_fs=16000)

# Clone the repo:
git clone https://github.com/alibaba/FunASR.git

# Install Conda:
wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh
sh Miniconda3-latest-Linux-x86_64.sh
conda create -n funasr python=3.7
conda activate funasr

# Install Pytorch (version >= 1.7.0):
conda install pytorch==1.7.0 torchvision==0.8.0 torchaudio==0.7.0 cudatoolkit=9.2 -c pytorch  # For more versions, please see https://pytorch.org/get-started/locally/

# Install other packages:
pip install --editable ./

pip install "modelscope[audio]" -f https://modelscope.oss-cn-beijing.aliyuncs.com/releases/repo.html

tree ./example_data/
./example_data/
├── dev
│   ├── text
│   └── wav.scp
├── test
│   ├── text
│   └── wav.scp
└── train
├── text
└── wav.scp

3 directories, 6 files

cat text
BAC009S0002W0122 而 对 楼市 成交 抑制 作用 最 大 的 限 购
BAC009S0002W0123 也 成为 地方 政府 的 眼中 钉

cat wav.scp
BAC009S0002W0122 /mnt/data/wav/train/S0002/BAC009S0002W0122.wav
BAC009S0002W0123 /mnt/data/wav/train/S0002/BAC009S0002W0123.wav

cd egs_modelscope/common

# compute fbank features
utils/compute_fbank.sh --cmd "utils/run.pl" --nj 32 --speed_perturb "1.0" \
example_data/train ${exp_dir}/exp/make_fbank/train ${fbankdir}/train
utils/fix_data_feat.sh ${fbankdir}/train

# apply low_frame_rate and cmvn
utils/apply_lfr_and_cmvn.sh --cmd "utils/run.pl" --nj 32 --lfr True --lfr-m 7 --lfr-n 6 \
${fbankdir}/train am.mvn ${exp_dir}/exp/make_fbank/train ${feat_train_dir}

# Text Tokenize
# 我爱reading->我 爱 read@@ ing
utils/text_tokenize.sh --cmd "utils/run.pl" --nj 32 ${fbankdir}/train seg_dict ${feat_train_dir}/log ${feat_train_dir}

# Dictionary Preparation
awk -v v=,vocab_size '{print $0v}' ${feat_train_dir}/text_shape > ${feat_train_dir}/text_shape.char
cp ${feat_train_dir}/speech_shape ${feat_train_dir}/text_shape ${feat_train_dir}/text_shape.char asr_stats_fbank_zh_char/train

# dev集参照如上train集对音频进行特征提取

modelname="speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
python modelscope_utils/download_model.py --model_name $modelname

python asr_train_paraformer.py \
--train_data_path_and_name_and_type ${feat_train_dir}/feats.scp,speech,kaldi_ark \
--train_data_path_and_name_and_type ${feat_train_dir}/text,text,text \
--valid_data_path_and_name_and_type ${feat_dev_dir}/feats.sc,speech,kaldi_ark \
--valid_data_path_and_name_and_type ${feat_dev_dir}/text,text,text \
--output_dir exp/model_dir \
--init_param base_model.pth \
--config finetune_config.yaml

cp finetuned_model.pth exp_dir/finetune_model_name.modelscope
cp -r ${HOME}/.cache/modelscope/hub/damo/pretrained_model_name/* exp_dir/

python -m funasr.bin.modelscope_infer \
--local_model_path exp_dir \
--wav_list example_data/dev/wav.scp \
--output_file logdir/text

# 配置modelscope_common_finetune.sh中参数
# dataset：数据路径，结构如example_data所示，dev/test可不配置，若无dev数据处理过程中可自动抽取训练集中1000句音频作为dev集
# tag：结果保存路径后缀

# 配置修改完成后，执行命令启动模型微调训练及推理:
sh modelscope_common_finetune.sh

cd egs_modelscope/speechio/paraformer

# 配置paraformer_large_infer.sh中参数
# ori_data:       # 测试数据路径
# data_dir:        # 数据处理路径
# exp_dir:        # 结果保存路径
# test_sets:       # 测试集名称
# use_lm=true      # 是否使用LM
# beam_size=10    # 设置beam_size
# lm_weight=0.15  # 设置lm_weight

# 测试集目录结构树，必须有trans.txt和wav.scp
tree SPEECHIO_ASR_ZH00001
SPEECHIO_ASR_ZH00001
├── metadata.tsv
├── trans.txt
├── wav
│   ├── e5oipIfM49I__20190201_CCTV_10.wav
│   └── e5oipIfM49I__20190201_CCTV_1.wav
│    └── ...
└── wav.scp


# 配置修改完成后，执行命令启动模型推理:
sh paraformer_large_infer.sh