jam134 commited on
Commit
cbc71d7
·
verified ·
1 Parent(s): d1a3893

Upload tools/asr\models\speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch\README.md with huggingface_hub

Browse files
tools/asr//models//speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch//README.md ADDED
@@ -0,0 +1,407 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ tasks:
3
+ - auto-speech-recognition
4
+ domain:
5
+ - audio
6
+ model-type:
7
+ - Non-autoregressive
8
+ frameworks:
9
+ - pytorch
10
+ backbone:
11
+ - transformer/conformer
12
+ metrics:
13
+ - CER
14
+ license: Apache License 2.0
15
+ language:
16
+ - cn
17
+ tags:
18
+ - FunASR
19
+ - Paraformer
20
+ - Alibaba
21
+ - INTERSPEECH 2022
22
+ datasets:
23
+ train:
24
+ - 60,000 hour industrial Mandarin task
25
+ test:
26
+ - AISHELL-1 dev/test
27
+ - AISHELL-2 dev_android/dev_ios/dev_mic/test_android/test_ios/test_mic
28
+ - WentSpeech dev/test_meeting/test_net
29
+ - SpeechIO TIOBE
30
+ - 60,000 hour industrial Mandarin task
31
+ indexing:
32
+ results:
33
+ - task:
34
+ name: Automatic Speech Recognition
35
+ dataset:
36
+ name: 60,000 hour industrial Mandarin task
37
+ type: audio # optional
38
+ args: 16k sampling rate, 8404 characters # optional
39
+ metrics:
40
+ - type: CER
41
+ value: 8.53% # float
42
+ description: greedy search, withou lm, avg.
43
+ args: default
44
+ - type: RTF
45
+ value: 0.0251 # float
46
+ description: GPU inference on V100
47
+ args: batch_size=1
48
+ widgets:
49
+ - task: auto-speech-recognition
50
+ inputs:
51
+ - type: audio
52
+ name: input
53
+ title: 音频
54
+ examples:
55
+ - name: 1
56
+ title: 示例1
57
+ inputs:
58
+ - name: input
59
+ data: https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav
60
+ inferencespec:
61
+ cpu: 8 #CPU数量
62
+ memory: 4096
63
+ finetune-support: True
64
+ ---
65
+
66
+
67
+ # Paraformer-large模型介绍
68
+
69
+ ## Highlights
70
+ - 热词版本:[Paraformer-large热词版模型](https://www.modelscope.cn/models/damo/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404/summary)支持热词定制功能,基于提供的热词列表进行激励增强,提升热词的召回率和准确率。
71
+ - 长音频版本:[Paraformer-large长音频模型](https://www.modelscope.cn/models/damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary),集成VAD、ASR、标点与时间戳功能,可直接对时长为数小时音频进行识别,并输出带标点文字与时间戳。
72
+
73
+ ## <strong>[FunASR开源项目介绍](https://github.com/alibaba-damo-academy/FunASR)</strong>
74
+ <strong>[FunASR](https://github.com/alibaba-damo-academy/FunASR)</strong>希望在语音识别的学术研究和工业应用之间架起一座桥梁。通过发布工业级语音识别模型的训练和微调,研究人员和开发人员可以更方便地进行语音识别模型的研究和生产,并推动语音识别生态的发展。让语音识别更有趣!
75
+
76
+ [**github仓库**](https://github.com/alibaba-damo-academy/FunASR)
77
+ | [**最新动态**](https://github.com/alibaba-damo-academy/FunASR#whats-new)
78
+ | [**环境安装**](https://github.com/alibaba-damo-academy/FunASR#installation)
79
+ | [**服务部署**](https://www.funasr.com)
80
+ | [**模型库**](https://github.com/alibaba-damo-academy/FunASR/tree/main/model_zoo)
81
+ | [**联系我们**](https://github.com/alibaba-damo-academy/FunASR#contact)
82
+
83
+
84
+ ## 模型原理介绍
85
+
86
+ Paraformer是达摩院语音团队提出的一种高效的非自回归端到端语音识别框架。本项目为Paraformer中文通用语音识别模型,采用工业级数万小时的标注音频进行模型训练,保证了模型的通用识别效果。模型可以被应用于语音输入法、语音导航、智能会议纪要等场景。
87
+
88
+ <p align="center">
89
+ <img src="fig/struct.png" alt="Paraformer模型结构" width="500" />
90
+
91
+
92
+ Paraformer模型结构如上图所示,由 Encoder、Predictor、Sampler、Decoder 与 Loss function 五部分组成。Encoder可以采用不同的网络结构,例如self-attention,conformer,SAN-M等。Predictor 为两层FFN,预测目标文字个数以及抽取目标文字对应的声学向量。Sampler 为无可学习参数模块,依据输入的声学向量和目标向量,生产含有语义的特征向量。Decoder 结构与自回归模型类似,为双向建模(自回归为单向建模)。Loss function 部分,除了交叉熵(CE)与 MWER 区分性优化目标,还包括了 Predictor 优化目标 MAE。
93
+
94
+
95
+ 其核心点主要有:
96
+ - Predictor 模块:基于 Continuous integrate-and-fire (CIF) 的 预测器 (Predictor) 来抽取目标文字对应的声学特征向量,可以更加准确的预测语音中目标文字个数。
97
+ - Sampler:通过采样,将声学特征向量与目标文字向量变换成含有语义信息的特征向量,配合双向的 Decoder 来增强模型对于上下文的建模能力。
98
+ - 基于负样本采样的 MWER 训练准则。
99
+
100
+ 更详细的细节见:
101
+ - 论文: [Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition](https://arxiv.org/abs/2206.08317)
102
+ - 论文解读:[Paraformer: 高识别率、高计算效率的单轮非自回归端到端语音识别模型](https://mp.weixin.qq.com/s/xQ87isj5_wxWiQs4qUXtVw)
103
+
104
+
105
+ ## 基于ModelScope进行推理
106
+
107
+ - 推理支持音频格式如下:
108
+ - wav文件路径,例如:data/test/audios/asr_example.wav
109
+ - pcm文件路径,例如:data/test/audios/asr_example.pcm
110
+ - wav文件url,例如:https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav
111
+ - wav二进制数据,格式bytes,例如:用户直接从文件里读出bytes数据或者是麦克风录出bytes数据。
112
+ - 已解析的audio音频,例如:audio, rate = soundfile.read("asr_example_zh.wav"),类型为numpy.ndarray或者torch.Tensor。
113
+ - wav.scp文件,需符合如下要求:
114
+
115
+ ```sh
116
+ cat wav.scp
117
+ asr_example1 data/test/audios/asr_example1.wav
118
+ asr_example2 data/test/audios/asr_example2.wav
119
+ ...
120
+ ```
121
+
122
+ - 若输入格式wav文件url,api调用方式可参考如下范例:
123
+
124
+ ```python
125
+ from modelscope.pipelines import pipeline
126
+ from modelscope.utils.constant import Tasks
127
+
128
+ inference_pipeline = pipeline(
129
+ task=Tasks.auto_speech_recognition,
130
+ model='iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch', model_revision="v2.0.4")
131
+
132
+ rec_result = inference_pipeline('https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav')
133
+ print(rec_result)
134
+ ```
135
+
136
+ - 输入音频为pcm格式,调用api时需要传入音频采样率参数audio_fs,例如:
137
+
138
+ ```python
139
+ rec_result = inference_pipeline('https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.pcm', audio_fs=16000)
140
+ ```
141
+
142
+ - 输入音频为wav格式,api调用方式可参考如下范例:
143
+
144
+ ```python
145
+ rec_result = inference_pipeline('asr_example_zh.wav')
146
+ ```
147
+
148
+ - 若输入格式为文件wav.scp(注:文件名需要以.scp结尾),可添加 output_dir 参数将识别结果写入文件中,api调用方式可参考如下范例:
149
+
150
+ ```python
151
+ inference_pipeline("wav.scp", output_dir='./output_dir')
152
+ ```
153
+ 识别结果输出路径结构如下:
154
+
155
+ ```sh
156
+ tree output_dir/
157
+ output_dir/
158
+ └── 1best_recog
159
+ ├── score
160
+ └── text
161
+
162
+ 1 directory, 3 files
163
+ ```
164
+ score:识别路径得分
165
+
166
+ text:语音识别结果文件
167
+
168
+
169
+ - 若输入音频为已解析的audio音频,api调用方式可参考如下范例:
170
+
171
+ ```python
172
+ import soundfile
173
+
174
+ waveform, sample_rate = soundfile.read("asr_example_zh.wav")
175
+ rec_result = inference_pipeline(waveform)
176
+ ```
177
+
178
+ - ASR、VAD、PUNC模型自由组合
179
+
180
+ 可根据使用需求对VAD和PUNC标点模型进行自由组合,使用方式如下:
181
+ ```python
182
+ inference_pipeline = pipeline(
183
+ task=Tasks.auto_speech_recognition,
184
+ model='iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch', model_revision="v2.0.4",
185
+ vad_model='iic/speech_fsmn_vad_zh-cn-16k-common-pytorch', vad_model_revision="v2.0.4",
186
+ punc_model='iic/punc_ct-transformer_zh-cn-common-vocab272727-pytorch', punc_model_revision="v2.0.3",
187
+ # spk_model="iic/speech_campplus_sv_zh-cn_16k-common",
188
+ # spk_model_revision="v2.0.2",
189
+ )
190
+ ```
191
+ 若不使用PUNC模型,可配置punc_model="",或不传入punc_model参数,如需加入LM模型,可增加配置lm_model='damo/speech_transformer_lm_zh-cn-common-vocab8404-pytorch',并设置lm_weight和beam_size参数。
192
+
193
+ ## 基于FunASR进行推理
194
+
195
+ 下面为快速上手教程,测试音频([中文](https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/vad_example.wav),[英文](https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_en.wav))
196
+
197
+ ### 可执行命令行
198
+ 在命令行终端执行:
199
+
200
+ ```shell
201
+ funasr +model=paraformer-zh +vad_model="fsmn-vad" +punc_model="ct-punc" +input=vad_example.wav
202
+ ```
203
+
204
+ 注:支持单条音频文件识别,也支持文件列表,列表为kaldi风格wav.scp:`wav_id wav_path`
205
+
206
+ ### python示例
207
+ #### 非实时语音识别
208
+ ```python
209
+ from funasr import AutoModel
210
+ # paraformer-zh is a multi-functional asr model
211
+ # use vad, punc, spk or not as you need
212
+ model = AutoModel(model="paraformer-zh", model_revision="v2.0.4",
213
+ vad_model="fsmn-vad", vad_model_revision="v2.0.4",
214
+ punc_model="ct-punc-c", punc_model_revision="v2.0.4",
215
+ # spk_model="cam++", spk_model_revision="v2.0.2",
216
+ )
217
+ res = model.generate(input=f"{model.model_path}/example/asr_example.wav",
218
+ batch_size_s=300,
219
+ hotword='魔搭')
220
+ print(res)
221
+ ```
222
+ 注:`model_hub`:表示模型仓库,`ms`为选择modelscope下载,`hf`为选择huggingface下载。
223
+
224
+ #### 实时语音识别
225
+
226
+ ```python
227
+ from funasr import AutoModel
228
+
229
+ chunk_size = [0, 10, 5] #[0, 10, 5] 600ms, [0, 8, 4] 480ms
230
+ encoder_chunk_look_back = 4 #number of chunks to lookback for encoder self-attention
231
+ decoder_chunk_look_back = 1 #number of encoder chunks to lookback for decoder cross-attention
232
+
233
+ model = AutoModel(model="paraformer-zh-streaming", model_revision="v2.0.4")
234
+
235
+ import soundfile
236
+ import os
237
+
238
+ wav_file = os.path.join(model.model_path, "example/asr_example.wav")
239
+ speech, sample_rate = soundfile.read(wav_file)
240
+ chunk_stride = chunk_size[1] * 960 # 600ms
241
+
242
+ cache = {}
243
+ total_chunk_num = int(len((speech)-1)/chunk_stride+1)
244
+ for i in range(total_chunk_num):
245
+ speech_chunk = speech[i*chunk_stride:(i+1)*chunk_stride]
246
+ is_final = i == total_chunk_num - 1
247
+ res = model.generate(input=speech_chunk, cache=cache, is_final=is_final, chunk_size=chunk_size, encoder_chunk_look_back=encoder_chunk_look_back, decoder_chunk_look_back=decoder_chunk_look_back)
248
+ print(res)
249
+ ```
250
+
251
+ 注:`chunk_size`为流式延时配置,`[0,10,5]`表示上屏实时出字粒度为`10*60=600ms`,未来信息为`5*60=300ms`。每次推理输入为`600ms`(采样点数为`16000*0.6=960`),输出为对应文字,最后一个语音片段输入需要设置`is_final=True`来强制输出最后一个字。
252
+
253
+ #### 语音端点检测(非实时)
254
+ ```python
255
+ from funasr import AutoModel
256
+
257
+ model = AutoModel(model="fsmn-vad", model_revision="v2.0.4")
258
+
259
+ wav_file = f"{model.model_path}/example/asr_example.wav"
260
+ res = model.generate(input=wav_file)
261
+ print(res)
262
+ ```
263
+
264
+ #### 语音端点检测(实时)
265
+ ```python
266
+ from funasr import AutoModel
267
+
268
+ chunk_size = 200 # ms
269
+ model = AutoModel(model="fsmn-vad", model_revision="v2.0.4")
270
+
271
+ import soundfile
272
+
273
+ wav_file = f"{model.model_path}/example/vad_example.wav"
274
+ speech, sample_rate = soundfile.read(wav_file)
275
+ chunk_stride = int(chunk_size * sample_rate / 1000)
276
+
277
+ cache = {}
278
+ total_chunk_num = int(len((speech)-1)/chunk_stride+1)
279
+ for i in range(total_chunk_num):
280
+ speech_chunk = speech[i*chunk_stride:(i+1)*chunk_stride]
281
+ is_final = i == total_chunk_num - 1
282
+ res = model.generate(input=speech_chunk, cache=cache, is_final=is_final, chunk_size=chunk_size)
283
+ if len(res[0]["value"]):
284
+ print(res)
285
+ ```
286
+
287
+ #### 标点恢复
288
+ ```python
289
+ from funasr import AutoModel
290
+
291
+ model = AutoModel(model="ct-punc", model_revision="v2.0.4")
292
+
293
+ res = model.generate(input="那今天的会就到这里吧 happy new year 明年见")
294
+ print(res)
295
+ ```
296
+
297
+ #### 时间戳预测
298
+ ```python
299
+ from funasr import AutoModel
300
+
301
+ model = AutoModel(model="fa-zh", model_revision="v2.0.4")
302
+
303
+ wav_file = f"{model.model_path}/example/asr_example.wav"
304
+ text_file = f"{model.model_path}/example/text.txt"
305
+ res = model.generate(input=(wav_file, text_file), data_type=("sound", "text"))
306
+ print(res)
307
+ ```
308
+
309
+ 更多详细用法([示例](https://github.com/alibaba-damo-academy/FunASR/tree/main/examples/industrial_data_pretraining))
310
+
311
+
312
+ ## 微调
313
+
314
+ 详细用法([示例](https://github.com/alibaba-damo-academy/FunASR/tree/main/examples/industrial_data_pretraining))
315
+
316
+
317
+ ## Benchmark
318
+ 结合大数据、大模型优化的Paraformer在一序列语音识别的benchmark上获得当前SOTA的效果,以下展示学术数据集AISHELL-1、AISHELL-2、WenetSpeech,公开评测项目SpeechIO TIOBE白盒测试场景的效果。在学术界常用的中文语音识别评测任务中,其表现远远超于目前公开发表论文中的结果,远好于单独封闭数据集上的模型。
319
+
320
+ ### AISHELL-1
321
+
322
+ | AISHELL-1 test | w/o LM | w/ LM |
323
+ |:------------------------------------------------:|:-------------------------------------:|:-------------------------------------:|
324
+ | <div style="width: 150pt">Espnet</div> | <div style="width: 150pt">4.90</div> | <div style="width: 150pt">4.70</div> |
325
+ | <div style="width: 150pt">Wenet</div> | <div style="width: 150pt">4.61</div> | <div style="width: 150pt">4.36</div> |
326
+ | <div style="width: 150pt">K2</div> | <div style="width: 150pt">-</div> | <div style="width: 150pt">4.26</div> |
327
+ | <div style="width: 150pt">Blockformer</div> | <div style="width: 150pt">4.29</div> | <div style="width: 150pt">4.05</div> |
328
+ | <div style="width: 150pt">Paraformer-large</div> | <div style="width: 150pt">1.95</div> | <div style="width: 150pt">1.68</div> |
329
+
330
+ ### AISHELL-2
331
+
332
+ | | dev_ios| test_android| test_ios|test_mic|
333
+ |:-------------------------------------------------:|:-------------------------------------:|:-------------------------------------:|:------------------------------------:|:------------------------------------:|
334
+ | <div style="width: 150pt">Espnet</div> | <div style="width: 70pt">5.40</div> |<div style="width: 70pt">6.10</div> |<div style="width: 70pt">5.70</div> |<div style="width: 70pt">6.10</div> |
335
+ | <div style="width: 150pt">WeNet</div> | <div style="width: 70pt">-</div> |<div style="width: 70pt">-</div> |<div style="width: 70pt">5.39</div> |<div style="width: 70pt">-</div> |
336
+ | <div style="width: 150pt">Paraformer-large</div> | <div style="width: 70pt">2.80</div> |<div style="width: 70pt">3.13</div> |<div style="width: 70pt">2.85</div> |<div style="width: 70pt">3.06</div> |
337
+
338
+
339
+ ### Wenetspeech
340
+
341
+ | | dev| test_meeting| test_net|
342
+ |:-------------------------------------------------:|:-------------------------------------:|:-------------------------------------:|:------------------------------------:|
343
+ | <div style="width: 150pt">Espnet</div> | <div style="width: 100pt">9.70</div> |<div style="width: 100pt">15.90</div> |<div style="width: 100pt">8.80</div> |
344
+ | <div style="width: 150pt">WeNet</div> | <div style="width: 100pt">8.60</div> |<div style="width: 100pt">17.34</div> |<div style="width: 100pt">9.26</div> |
345
+ | <div style="width: 150pt">K2</div> | <div style="width: 100pt">7.76</div> |<div style="width: 100pt">13.41</div> |<div style="width: 100pt">8.71</div> |
346
+ | <div style="width: 150pt">Paraformer-large</div> | <div style="width: 100pt">3.57</div> |<div style="width: 100pt">6.97</div> |<div style="width: 100pt">6.74</div> |
347
+
348
+ ### SpeechIO TIOBE
349
+
350
+ Paraformer-large模型结合Transformer-LM模型做shallow fusion,在公开评测项目SpeechIO TIOBE白盒测试场景上获得当前SOTA的效果,目前[Transformer-LM模型](https://modelscope.cn/models/damo/speech_transformer_lm_zh-cn-common-vocab8404-pytorch/summary)已在ModelScope上开源,以下展示SpeechIO TIOBE白盒测试场景without LM、with Transformer-LM的效果:
351
+
352
+ - Decode config w/o LM:
353
+ - Decode without LM
354
+ - Beam size: 1
355
+ - Decode config w/ LM:
356
+ - Decode with [Transformer-LM](https://modelscope.cn/models/damo/speech_transformer_lm_zh-cn-common-vocab8404-pytorch/summary)
357
+ - Beam size: 10
358
+ - LM weight: 0.15
359
+
360
+ | testset | w/o LM | w/ LM |
361
+ |:------------------:|:----:|:----:|
362
+ |<div style="width: 200pt">SPEECHIO_ASR_ZH00001</div>| <div style="width: 150pt">0.49</div> | <div style="width: 150pt">0.35</div> |
363
+ |<div style="width: 200pt">SPEECHIO_ASR_ZH00002</div>| <div style="width: 150pt">3.23</div> | <div style="width: 150pt">2.86</div> |
364
+ |<div style="width: 200pt">SPEECHIO_ASR_ZH00003</div>| <div style="width: 150pt">1.13</div> | <div style="width: 150pt">0.80</div> |
365
+ |<div style="width: 200pt">SPEECHIO_ASR_ZH00004</div>| <div style="width: 150pt">1.33</div> | <div style="width: 150pt">1.10</div> |
366
+ |<div style="width: 200pt">SPEECHIO_ASR_ZH00005</div>| <div style="width: 150pt">1.41</div> | <div style="width: 150pt">1.18</div> |
367
+ |<div style="width: 200pt">SPEECHIO_ASR_ZH00006</div>| <div style="width: 150pt">5.25</div> | <div style="width: 150pt">4.85</div> |
368
+ |<div style="width: 200pt">SPEECHIO_ASR_ZH00007</div>| <div style="width: 150pt">5.51</div> | <div style="width: 150pt">4.97</div> |
369
+ |<div style="width: 200pt">SPEECHIO_ASR_ZH00008</div>| <div style="width: 150pt">3.69</div> | <div style="width: 150pt">3.18</div> |
370
+ |<div style="width: 200pt">SPEECHIO_ASR_ZH00009</div>| <div style="width: 150pt">3.02</div> | <div style="width: 150pt">2.78</div> |
371
+ |<div style="width: 200pt">SPEECHIO_ASR_ZH000010</div>| <div style="width: 150pt">3.35</div> | <div style="width: 150pt">2.99</div> |
372
+ |<div style="width: 200pt">SPEECHIO_ASR_ZH000011</div>| <div style="width: 150pt">1.54</div> | <div style="width: 150pt">1.25</div> |
373
+ |<div style="width: 200pt">SPEECHIO_ASR_ZH000012</div>| <div style="width: 150pt">2.06</div> | <div style="width: 150pt">1.68</div> |
374
+ |<div style="width: 200pt">SPEECHIO_ASR_ZH000013</div>| <div style="width: 150pt">2.57</div> | <div style="width: 150pt">2.25</div> |
375
+ |<div style="width: 200pt">SPEECHIO_ASR_ZH000014</div>| <div style="width: 150pt">3.86</div> | <div style="width: 150pt">3.08</div> |
376
+ |<div style="width: 200pt">SPEECHIO_ASR_ZH000015</div>| <div style="width: 150pt">3.34</div> | <div style="width: 150pt">2.67</div> |
377
+
378
+
379
+ ## 使用方式以及适用范围
380
+
381
+ 运行范围
382
+ - 支持Linux-x86_64、Mac和Windows运行。
383
+
384
+ 使用方式
385
+ - 直接推理:可以直接对输入音频进行解码,输出目标文字。
386
+ - 微调:加载训练好的模型,采用私有或者开源数据进行模型训练。
387
+
388
+ 使用范围与目标场景
389
+ - 适合与离线语音识别场景,如录音文件转写,配合GPU推理效果更加,推荐输入语音时长在20s以下,若想解码长音频,推荐使用[Paraformer-large长音频模型](https://www.modelscope.cn/models/damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary),集成VAD、ASR、标点与时间戳功能,可直接对时长为数小时音频进行识别,并输出带标点文字与时间戳。
390
+
391
+
392
+ ## 模型局限性以及可能的偏差
393
+
394
+ 考虑到特征提取流程和工具以及训练工具差异,会对CER的数据带来一定的差异(<0.1%),推理GPU环境差异导致的RTF数值差异。
395
+
396
+
397
+
398
+ ## 相关论文以及引用信息
399
+
400
+ ```BibTeX
401
+ @inproceedings{gao2022paraformer,
402
+ title={Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition},
403
+ author={Gao, Zhifu and Zhang, Shiliang and McLoughlin, Ian and Yan, Zhijie},
404
+ booktitle={INTERSPEECH},
405
+ year={2022}
406
+ }
407
+ ```