Experiment on creating a new dataset audio+text about deepspeech-italian-model HOT 3 OPEN

I'm starting same test of long audio segmentation, considering the speaker's voice activity.
On this fork: https://github.com/eziolotta/rVADfast

But i have same problem with quality of audio output...

from deepspeech-italian-model.

First experiment of segmentation of short audio, using rVADfast and an algorithm that analyze segments found by rVAD to generate a new sequence of speech segments.
rVAD (and same other) tend to cut last bit signal of a speech segment.
Code and other tests yet to be published.

Input Clip : 644_2532_000000.wav - 15 second - (MLS Dataset)
Output : 5 Speech Segments (wav files)

test_segmentation_short_audio.zip

i try to extend algo to long audio (maybe hour, try Public Podcast )

from deepspeech-italian-model.

Continuing the experiments with rVADFast, I was able to segment one random Podcast of Emilia Romagna Region

https://ambiente.regione.emilia-romagna.it/it/gallery/video/i-video-di-ermesambiente/convegno-inspire/stefano-olivucci-regione-emilia-romagna

Obtaining 143 segments with a duration from a minimum of 2 seconds to a maximum of 2 minutes.
Execution time for this process was approximately 1.5 hours

Audios are without transcription, so in this case an automatic transcription and human validation must be applied.

Unfortunately, other Speakers are also involved in podcasts, and some time words are not clear, check is required during validation. There is no background noise in Podcasts and the audio is clean.

Other Podcast here
Licence: Creative Commons Attribution 4.0

Output Dataset of My experiment can be downloaded here:
http://t.ly/xHHL

from deepspeech-italian-model.