twerkmeister/ilid

automatic spoken language identification

ilid's Issues

IndexError in evaluation/predict.py (classifier.py)

Hello twerkmeister, sorry if is trivial, is my first experience with caffe, I succede to preprocessing the audios, trained it with Berlin_net model, but at the evaluation step I get:

Traceback (most recent call last):
File "predict.py", line 47, in
predict(args.input, args.proto, args.model, args.output)
File "predict.py", line 20, in predict
raw_scale=255 # convert 0..255 values into range 0..1
File "/home/sylvain/caffe/python/caffe/classifier.py", line 29, in init
in_ = self.inputs[0]
IndexError: list index out of range

I put the complete log in attachment... If someone can help me. Thanks
output.log

Is there a reported accuracy and latency to predict a wav file

would someone tell me the accuracy of iLID and latency to decode a wav file of 10s?
I want to have the reference.
Regards,
Luke

what does the training data look like?

Hi, the link to the training data repo is dead, could you fix that? I cant run the train.py while I dont know what is in the trainingData.csv

Are there other advices to preprocess audio files?

Hi, twerkmeister. I'm following your excellent work 'iLID' recently. The approach shows good performance when tested on the dataset consisting of lots of clean audios. However, when tested on the audios recorded in natural scenes, it doesn't perform as well as before. In your project, I've seen the loudness normalization operation. Are there other advices to preprocess the audio to make it more clean?

many thanks.

ValueError: need more than 0 values to unpack

When i ran the train.py in the tensorflow, i got this issues. What this mean? Do i need some other things to be done, when I start to run the train program?

umlaut

Hi Thomas, one question: did you get any insight regarding to your question here? mozilla/TTS#232 (comment)
Do you mind if you share with us what solution you found for espeak-ng for the german umlauts? (if you found it)

Thank you!

How are silences treated?

I see you've applied sliding windows to downloaded audios. How did you manage silences or non speech windos in those audios?