Please read and document yourself on machine learning.
what do you mean with dataset ? my audio and text files ? i will give you any dataset, bu i dont get what do you mean with datasets ?
Yes. Also, why don’t you give a dev set. You can’t train seriously without
https://drive.google.com/open?id=1BDWNbsqSnrYw1i3MqfO4r0Is312DgfWu Here is datasets csv
https://drive.google.com/open?id=1X5wHUui0BPDRaxnOOAMtOTn5t3UFfUBv here is my audios
what parameters needed ? and what proportions ? epochs, n_hidden, learning_rate, drop_out_rate ? Where can i find informations about these paramateres ?
You misunderstood me. I don’t have the time to examinate your dataset. Please explain it.
Have you read the documentation ? They are documented. You also really need to get some experience on machine learning.
@Akmal_Nodirov This is really getting tiring now, though. You have not explained why you don’t have a dev set on your training.
Brother, i dont really understand you, how to explain my datasets, i ll try:
i have 93 phrases, not single words. Phrases like “Hello this is me”, “My name is someone” and so on.i I gave 10 of them to test, and another ten 10 dev set, leaving 73 for train like:
–train_batch_size 73 *
** --test_batch_size 10 *
** --dev_batch_size 10 **
** --n_hidden 100 **
** --epochs 200 **
and i have .wav audio file for every phrase
this is what youre asking of datasets ?
Ok, so very very small, hand-crafted dataset. Now we progress. How much audio does that make, in time? One hour?
What kind of use-case do you target ?
the audios are 10 minutes long. All audios are 10 minutes
Ok, so you have roughly 15 hours ? 10*93 makes 930 mins.
no, 93 phrase lasts 10 minutes)
You can’t expect to train anything with just 10 minutes. Again, what use-case do you target ?
Im uzbek, and there arent any fine uzbek SST, im goin to develop it) i mean my company, my city can use it after developing it. no commercial user for now
You need more than 10 minutes for generic-purpose STT. Are you contributing to Common Voice for Uzbek? That’s the best course of action at the moment.
Once you get a few dozen of hours, you can try and start building something with transfer-learning.
Ok thank you brother, i will try
1 Like
one question, can i repeate phrases ? for example above 93 phrases are only 13 different phrases. I just have repeated them. For example i have this phrase:
“Plastik yoqolsa bankka borib ochtiring”. This phrase repeated 7 times. and also 7 different audio with different tempo
We don’t have enough feedback yet on the behavior of DeepSpeech on that, it might depend on your dataset as well as your language. Repeated sentences (i.e. same sentence spoken by different people, so close to your case) seems to improve a bit the model. But we are far from having a definitive answer on that, so I would suggest to check that cautiously.
@Akmal_Nodirov A good way to start is to retrain something that is well known. Have you seen this video?
And for the number of samples. Typically, you will need about 10.000 samples of length 4-8 seconds to get somewhat good result for general language understanding.
Maybe you can get them here?