aishell example

游雁

2024-02-19 ce4235b1c83752841a8c3506da28f08607b56361

aishell example

 examples/aishell/paraformer/conf/train_asr_paraformer_conformer_12e_6d_2048_256.yaml

@@ -99,7 +99,10 @@
    max_token_length: 2048 # filter samples if source_token_len+target_token_len > max_token_length,
    buffer_size: 1024
    shuffle: True
    num_workers: 0
    num_workers: 4
    preprocessor_speech: SpeechPreprocessSpeedPerturb
    preprocessor_speech_conf:
      speed_perturb: [0.9, 1.0, 1.1]

tokenizer: CharTokenizer
tokenizer_conf:

 examples/aishell/paraformer/run.sh

@@ -1,13 +1,8 @@
#!/usr/bin/env bash

workspace=`pwd`

# machines configuration

CUDA_VISIBLE_DEVICES="0,1"
gpu_num=2
gpu_inference=true  # Whether to perform gpu decoding, set false for cpu decoding
# for gpu decoding, inference_nj=ngpu*njob; for cpu decoding, inference_nj=njob
njob=1

# general configuration
feats_dir="../DATA" #feature output dictionary
@@ -18,7 +13,11 @@
stop_stage=5

# feature configuration
nj=64
nj=32

inference_device="cuda" #"cpu"
inference_checkpoint="model.pt"
inference_scp="wav.scp"

# data
raw_data=../raw_data
@@ -26,6 +25,7 @@

# exp tag
tag="exp1"
workspace=`pwd`

. utils/parse_options.sh || exit 1;

@@ -41,11 +41,6 @@

config=train_asr_paraformer_conformer_12e_6d_2048_256.yaml
model_dir="baseline_$(basename "${config}" .yaml)_${lang}_${token_type}_${tag}"

inference_device="cuda" #"cpu"
inference_checkpoint="model.pt"
inference_scp="wav.scp"



if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
@@ -112,6 +107,8 @@
  mkdir -p ${exp_dir}/exp/${model_dir}
  log_file="${exp_dir}/exp/${model_dir}/train.log.txt"
  echo "log_file: ${log_file}"

  gpu_num=$(echo CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
  torchrun \
  --nnodes 1 \
  --nproc_per_node ${gpu_num} \

 funasr/datasets/audio_datasets/preprocessor.py

@@ -41,43 +41,9 @@
                 **kwargs):
        super().__init__()
        
        self.seg_dict = None
        if seg_dict is not None:
            self.seg_dict = {}
            with open(seg_dict, "r", encoding="utf8") as f:
                lines = f.readlines()
            for line in lines:
                s = line.strip().split()
                key = s[0]
                value = s[1:]
                self.seg_dict[key] = " ".join(value)
        self.text_cleaner = TextCleaner(text_cleaner)
        self.split_with_space = split_with_space
    
    def forward(self, text, **kwargs):
        if self.seg_dict is not None:
            text = self.text_cleaner(text)
            if self.split_with_space:
                tokens = text.strip().split(" ")
                if self.seg_dict is not None:
                    text = seg_tokenize(tokens, self.seg_dict)

        return text

def seg_tokenize(txt, seg_dict):
    pattern = re.compile(r'^[\u4E00-\u9FA50-9]+$')
    out_txt = ""
    for word in txt:
        word = word.lower()
        if word in seg_dict:
            out_txt += seg_dict[word] + " "
        else:
            if pattern.match(word):
                for char in word:
                    if char in seg_dict:
                        out_txt += seg_dict[char] + " "
                    else:
                        out_txt += "<unk>" + " "
            else:
                out_txt += "<unk>" + " "
    return out_txt.strip().split()

 funasr/frontends/wav_frontend.py

@@ -32,7 +32,6 @@
                rescale_line = line_item[3:(len(line_item) - 1)]
                vars_list = list(rescale_line)
                continue
    import pdb;pdb.set_trace()
    means = np.array(means_list).astype(np.float32)
    vars = np.array(vars_list).astype(np.float32)
    cmvn = np.array([means, vars])

 funasr/tokenizer/char_tokenizer.py

@@ -3,6 +3,7 @@
from typing import List
from typing import Union
import warnings
import re

from funasr.tokenizer.abs_tokenizer import BaseTokenizer
from funasr.register import tables
@@ -14,6 +15,8 @@
        non_linguistic_symbols: Union[Path, str, Iterable[str]] = None,
        space_symbol: str = "<space>",
        remove_non_linguistic_symbols: bool = False,
        split_with_space: bool = False,
        seg_dict: str = None,
        **kwargs,
    ):
        super().__init__(**kwargs)
@@ -31,6 +34,11 @@
        else:
            self.non_linguistic_symbols = set(non_linguistic_symbols)
        self.remove_non_linguistic_symbols = remove_non_linguistic_symbols
        self.split_with_space = split_with_space
        self.seg_dict = None
        if seg_dict is not None:
            self.seg_dict = load_seg_dict(seg_dict)
	

    def __repr__(self):
        return (
@@ -41,6 +49,12 @@
        )

    def text2tokens(self, line: Union[str, list]) -> List[str]:
		
        if self.split_with_space:
            tokens = line.strip().split(" ")
            if self.seg_dict is not None:
                tokens = seg_tokenize(tokens, self.seg_dict)
        else:
        tokens = []
        while len(line) != 0:
            for w in self.non_linguistic_symbols:
@@ -60,3 +74,34 @@
    def tokens2text(self, tokens: Iterable[str]) -> str:
        tokens = [t if t != self.space_symbol else " " for t in tokens]
        return "".join(tokens)


def load_seg_dict(seg_dict_file):
    seg_dict = {}
    assert isinstance(seg_dict_file, str)
    with open(seg_dict_file, "r", encoding="utf8") as f:
        lines = f.readlines()
        for line in lines:
            s = line.strip().split()
            key = s[0]
            value = s[1:]
            seg_dict[key] = " ".join(value)
    return seg_dict

def seg_tokenize(txt, seg_dict):
    pattern = re.compile(r'^[\u4E00-\u9FA50-9]+$')
    out_txt = ""
    for word in txt:
        word = word.lower()
        if word in seg_dict:
            out_txt += seg_dict[word] + " "
        else:
            if pattern.match(word):
                for char in word:
                    if char in seg_dict:
                        out_txt += seg_dict[char] + " "
                    else:
                        out_txt += "<unk>" + " "
            else:
                out_txt += "<unk>" + " "
    return out_txt.strip().split()