From 4ace5a95b052d338947fc88809a440ccd55cf6b4 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 16 十一月 2023 16:39:52 +0800
Subject: [PATCH] funasr pages
---
funasr/datasets/large_datasets/dataset.py | 41 ++++++++++++++++++++++++++---------------
1 files changed, 26 insertions(+), 15 deletions(-)
diff --git a/funasr/datasets/large_datasets/dataset.py b/funasr/datasets/large_datasets/dataset.py
index 68b63e1..adfe4f6 100644
--- a/funasr/datasets/large_datasets/dataset.py
+++ b/funasr/datasets/large_datasets/dataset.py
@@ -6,6 +6,8 @@
import torch
import torch.distributed as dist
import torchaudio
+import numpy as np
+import soundfile
from kaldiio import ReadHelper
from torch.utils.data import IterableDataset
@@ -106,7 +108,7 @@
ark_reader = ReadHelper('ark:{}'.format(data_file))
reader_list.append(ark_reader)
elif data_type == "text" or data_type == "sound" or data_type == 'text_hotword':
- text_reader = open(data_file, "r")
+ text_reader = open(data_file, "r", encoding="utf-8")
reader_list.append(text_reader)
elif data_type == "none":
continue
@@ -123,7 +125,14 @@
sample_dict["key"] = key
elif data_type == "sound":
key, path = item.strip().split()
- waveform, sampling_rate = torchaudio.load(path)
+ try:
+ waveform, sampling_rate = torchaudio.load(path)
+ except:
+ waveform, sampling_rate = soundfile.read(path, dtype='float32')
+ if waveform.ndim == 2:
+ waveform = waveform[:, 0]
+ waveform = np.expand_dims(waveform, axis=0)
+ waveform = torch.tensor(waveform)
if self.frontend_conf is not None:
if sampling_rate != self.frontend_conf["fs"]:
waveform = torchaudio.transforms.Resample(orig_freq=sampling_rate,
@@ -193,21 +202,23 @@
data_types = conf.get("data_types", "kaldi_ark,text")
pre_hwfile = conf.get("pre_hwlist", None)
- pre_prob = conf.get("pre_prob", 0) # unused yet
+ # pre_prob = conf.get("pre_prob", 0) # unused yet
+ if pre_hwfile is not None:
+ pre_hwlist = []
+ with open(pre_hwfile, 'r', encoding="utf-8") as fin:
+ for line in fin.readlines():
+ pre_hwlist.append(line.strip())
+ else:
+ pre_hwlist = None
hw_config = {"sample_rate": conf.get("sample_rate", 0.6),
"double_rate": conf.get("double_rate", 0.1),
"hotword_min_length": conf.get("hotword_min_length", 2),
"hotword_max_length": conf.get("hotword_max_length", 8),
- "pre_prob": conf.get("pre_prob", 0.0)}
+ "pre_prob": conf.get("pre_prob", 0.0),
+ "pre_hwlist": pre_hwlist}
- if pre_hwfile is not None:
- pre_hwlist = []
- with open(pre_hwfile, 'r') as fin:
- for line in fin.readlines():
- pre_hwlist.append(line.strip())
- else:
- pre_hwlist = None
+
dataset = AudioDataset(scp_lists,
data_names,
@@ -218,15 +229,15 @@
mode=mode,
)
- filter_conf = conf.get('filter_conf', {})
- filter_fn = partial(filter, **filter_conf)
- dataset = FilterIterDataPipe(dataset, fn=filter_fn)
-
if "text" in data_names:
vocab = {'vocab': dict, 'seg_dict': seg_dict, 'punc_dict': punc_dict, 'bpe_tokenizer': bpe_tokenizer, 'hw_config': hw_config}
tokenize_fn = partial(tokenize, **vocab)
dataset = MapperIterDataPipe(dataset, fn=tokenize_fn)
+ filter_conf = conf.get('filter_conf', {})
+ filter_fn = partial(filter, **filter_conf)
+ dataset = FilterIterDataPipe(dataset, fn=filter_fn)
+
if shuffle:
buffer_conf = conf.get('shuffle_conf', {})
buffer_size = buffer_conf['shuffle_size']
--
Gitblit v1.9.1