Test on data_for_model/test.csv - WER: 0.989899, CER: 0.888451, loss: 201.759644
--------------------------------------------------------------------------------
WER: 1.000000, CER: 0.882353, loss: 74.156654
- src: "this is demo"
- res: "i see"
--------------------------------------------------------------------------------
WER: 1.000000, CER: 0.888889, loss: 79.565163
- src: "compass computrace"
- res: "i see"
--------------------------------------------------------------------------------
WER: 1.000000, CER: 0.800000, loss: 87.420784
- src: "quality of food "
- res: "i see"
--------------------------------------------------------------------------------
WER: 1.000000, CER: 0.807692, loss: 103.913086
- src: "open the lock"
- res: "i see"
--------------------------------------------------------------------------------
WER: 1.000000, CER: 0.826087, loss: 109.038353
- src: "clear understanding demo"
- res: "i see"
--------------------------------------------------------------------------------
same result are coming any one please help me?
I am unable to improve model accuracy?
I know loss is high but res
are same for all src
?
I am doing any mistake please correct me?
lissyx
((slow to reply) [NOT PROVIDING SUPPORT])
February 21, 2020, 12:19pm
2
Who knows. We don’t know anything about your usage. Don’t expect anybody to be able to help you without sharing details.
Sorry @lissyx
About dataset I have only 450 wav file
train-317
dev- 84
other in test
train using below command:
python DeepSpeech.py
--train_files data_for_model/train.csv
--dev_files data_for_model/dev.csv
--test_files data_for_model/test.csv
--epochs 30
--learning_rate 0.0001
--export_dir ckexport/
--checkpoint_dir ckpoint/
after i am getting above result?
lissyx
((slow to reply) [NOT PROVIDING SUPPORT])
February 21, 2020, 12:56pm
4
How did you produced all of thoses ? What version are you working on ?
I am using DeepSpeech v0.5.1
wav recording with CHANNELS = 1
RATE = 16000
I have question in all recording starting with open
eg.
100.wav 67363 open this message
200.wav 67263 open all pages
like in all recording is it good? Sir
lissyx
((slow to reply) [NOT PROVIDING SUPPORT])
February 21, 2020, 1:08pm
6
I have no idea what you are doing.
@lissyx Sir I am having multiple recordings starting or containing similar word for
ex open mail
open phone ,
open calender
folder open etc
occurrence of open
is multiple time
my all data like that is it good ?
lissyx
((slow to reply) [NOT PROVIDING SUPPORT])
February 21, 2020, 1:20pm
8
I’ve asked you questions, you have not answered them and you still hide a lot of details of your context. I can’t help you if you don’t want help.
I am not able to train good model
I want to train a model with some command like above.
question is if similar words occurs in multiple time recording so model will train perfectly or not ?
othiele
(Olaf Thiele)
February 21, 2020, 4:17pm
10
Modern artificial intelligence relies on tons and tons of data. So if you had 30.000 files, you should get good results, I doubt that you get that with just 300.
You need more data.
@othiele and @lissyx Thanks for help