TUTORIAL : How I trained a specific french model to control my robot

DeepSpeech
#148

@elpimous_robot as you said for

TRIE CREATION :

you use

alphabet.txt
lm.binary
vacab.txt

and you create trie right?

is it required to use vocab.txt for trie creation ?

also one more question

for creation vocab.txt suppose in my wave transcripts repeated

eg:
1000.wav 23093 hello good morning [male voice]
2000.wav 32424 hello good morning [female voice]

so capy both sentence or not?

#149

Hello

Yes, and no !

For trie creation, you need textual sentences, to work with probabilities, accuracy…

Vocab.txt doesn’t need multiple times the same sentence

hope to help.
Vincent

#150

@elpimous_robot Thanks for help

1 Like
#151

Ready to help again my friend

#152

@elpimous_robot :blush:

#153

@elpimous_robot can you help me below issue

@victornoriega7 he is also help me in this issue

#154

hello deep_learning:
have you trained a korean model successly? i am training a korean model, but i meet much troubles ,could yuu please share some expirement with me ? how do you train it? thanks !

#155

Hello!
The way you explained the steps of training the model is amazing, I believe you helped a lot of people with this post, including me. I am a newbie regarding this whole concept and I’m stuck at the very first point of implementing a model for my native language. The csv file containing the wav_name, wav_size, transcript was done manually or by using a script that gathers these data? I have around 1500 .wav and I’m a little concerned about creating the csv manually (checking each .wav the size, etc). If there is a script that someone created in order to avoid all this struggle and would like to share…Thank you! :smiley:

#156

I did something like…

for i in $(ls *.wav); do
  fs=$(stat --printf="%s" $i)
  ts=$(cat $i.txt) ## assuming transcript for the wav file is file.wav.txt
  echo "$i,$fs,$ts" >> mydata.csv
done
1 Like
#158

Training on cpu?

If you have early stop on, it will stop training when it’s not seeing loss decrease significantly. As you’re on epoch 0 still you might need to adjust your settings a bit more or, if you’re cpu training, train on a gpu.

1 Like
#159

I was looking at mozilla voice-corpus-tool and I can’t see all of the effects that have been listed above by @elpimous_robot, the only ones that I can see on the github repo as well as when I run ‘help’ are:

Effects:

compr
Distortion by mp3 compression
Arguments:
kbit: int - Virtual bandwidth in kBit/s

rate
Resampling to different sample rate
Arguments:
rate: int - Sample rate to apply

augment [-times ] [-gain ]
Augment samples of current buffer with noise
Arguments:
source: string - CSV file with samples to augment onto current sample buffer
Options:
-times: int - How often to apply the augmentation source to the sample buffer
-gain: float - How much gain (in dB) to apply to augmentation audio before overlaying onto buffer samples

#160

How to create this file libctc_decoder_with_kenlm.so ?

I have only lm.binary and trie file created using KenLM

I have seen you commands generate try with is have 5 parameter but now it won’t accept 5 parameter it required only 4 parameter vocab.txt not accept

Please help in that

#161

Please, make an effort and understand this tutorial was contributed by @elpimous_robot a long time ago, and that the project moved since, and a lot of the instructions are deprecated.

#162

can we use this method to prepare data and train deepspeech v0.6.0?

#163

Hello i have a question, i have a three type of files, train, dev and test, and the words and audio within these must be same ? i mean, dev files and test files must include in train files ?

#164

Please read the guidelines, it says explicitly don’t hijack old threads …

#165

where is that guidlines ?

#166

Really? I posted them with your first thread, that you abandoned to hijack this one, but here you go:

https://discourse.mozilla.org/t/what-and-how-to-report-if-you-need-support/62071/2

Please stop spamming other threads, do your homework and we are happy to help.

#167

hi, i have a problem,
FATAL Flags parsing error: flag --alphabet_config_path=/content/gdrive/My\ drive/deepSpeech/alphabet_rus.txt: The file pointed to by --alphabet_config_path must exist and be readable.
Pass --helpshort or --helpfull to see help on flags.

Here is my datasets:

#168

!python3 DeepSpeech.py \

--drop_source_layers 1 \

--alphabet_config_path "/content/gdrive/My\ drive/deepSpeech/alphabet_rus.txt" \

--save_checkpoint_dir /content/gdrive/My\ drive/deepSpeech/savecheckpoint \

--load_checkpoint_dir /content/gdrive/My\ drive/deepSpeech/loadcheckpoint \

--train_files   /content/gdrive/My\ drive/deepSpeech/train/train.csv \

--dev_files   /content/gdrive/My\ drive/deepSpeech/dev/dev.csv \

--test_files  /content/gdrive/My\ drive/deepSpeech/test/test.csv \