Specification of audio format files

linto-ai / linto-desktoptools-hmg

GUI Tool to create, manage and test Keyword Spotting models using TF 2.0

GNU Affero General Public License v3.0

12 stars 2 forks source link

Specification of audio format files #5

Closed ssosas closed 3 years ago

ssosas commented 3 years ago

Dear all, Could you be more specific about the format of audio files required?

extension (MP3, WAV, etc)??
recommended format (frequency sampling, duration, etc), what have you used for LinTO generation?
Does the duration need to be exactly the same? Are some millis of difference authorized?
may I have access to some examples?

thanks a lot! Sergio

Lokhozt commented 3 years ago

Hello,

For now only .wav 16 kHz mono 2B files are supported.
Duration can vary slightly, longer samples are trimmed, shorter samples are ignored in this version.
We provide our data there.

Note that the current available release is not up to date and the new version is still a work in progress.

Cheers

JulianGerhard21 commented 3 years ago

Hi @Lokhozt ,

thanks for clarification. I am currently running it on Ubuntu 20.04 and I am providing samples in the form of:

{"file_name": "u_xxx.wav", "label": "test", "attr": null}

Is this correct? Because if I do so and want to evaluate the data with Analyze Samples, your tool tells me that several files couldnt be found and crashes. After debugging those files, I found out, that the tool seemed to mix up the labelled files meaning that a sample containing the hotword was used as a sample for the non-hotword and hence couldnt be found in the specified directory.

I am a bit desparate because I tried a lot of situations/fixes but nothing seem to work. Would it be a problem to send you the files to simply train a model for me?

Kind regards Julian

Lokhozt commented 3 years ago

Hello Julian, Sorry for the delay i'm deep into 1.0 code and i missed the notification. Do you have a console trace of the error ?

JulianGerhard21 commented 3 years ago

Hi @Lokhozt ,

unfortunately, the tool can't be restarted now. I have attached the audio files in a zip, both the labelled ones and the noise. It would be a huge favor if you could quickly train a model with them... shouldn't last long. I'll wait until the release of 1.0 then.

audio_files.zip

I simply want to test something so a perfect quality is absolutely not necessary... I hope you are able to help!

The error was a bit weird. Basically the tool told me that it could not find a bunch of files. The weird thing is, that I added them previously. After debugging the error (no stacktrace, just the error) I realised that the connection between the label and the actual filepath was mixed up.

Kind regards and many thanks Julian

Lokhozt commented 3 years ago

Hello Julian, I have trouble with your samples, cannot open them properly. Cannot play them neither.