Understanding Audio Signals for Machine Learning 

Thank you for taking the time :)

Thanks!

Thank you Joel!

That's an incredibly deep rabbit hole :)

Thank you! I intend to cover MFCCs quite in depth. Stay tuned :)

You're welcome!

Thanks!

@@ValerioVelardoTheSoundofAI hey can please make a video of converting one instrumental sound to another or one person vocal to another in python.thanx in advance

Thanks!

Thank you! Unfortunately, there aren't many resources about AI audio. I'm currently writing a book on the topic. A book with "traditional" DSP approaches is "Fundamentals of Music Processing" www.springer.com/gp/book/9783319219448

@@ValerioVelardoTheSoundofAI I'll buy yours when it is ready!

@@JogosEtudoMais thanks!

I guess he intended to say 5.047MB and made a typing error in the slide which he read out.

A bit can either be 1 or 0 (On / Off) A byte is a group of 8 bits. So if there are 8 bits, where each bit can only be 0 or 1, there are a total of 2^8 = 256 different values that a byte can represent. So let's consider bit depth * sampling rate = 16 * 44,100 = 705,600 bits per second. There are 1,048,576 bits in megabits (mega means 1 million, but the closest binary representation of that is 2^20 = 1,048,576). So 705,600 / 1,048,576 = 0.6729 megabits per second. Remember that 8 bits = 1 byte? 8 megabits is also 1 megabyte. So we just need to divide by 8. 0.6729/8 = 0.0841 megabytes per second = 5.0468 megabytes per minute. I'm assuming it's a typo in the video.

@@johnnyvishnevskiy8090 You're a freaking legend, thank you so much. This all makes sense and now I'm gonna read it like 1,048 times. lol

@@Drew_7 np!

Yes, this great book: - Music Similarity and Retrieval www.springer.com/gp/book/9783662497203

@@ValerioVelardoTheSoundofAI Thanks for the suggestion.

I would stick with the files at 8KHz, which has the advantage of resulting in lighter data.

@@ValerioVelardoTheSoundofAI I believe that you are saying, it is to keep 8000 hz for audio feature extraction. Whether it is wrong/some effects will be there when we have tuned sampling rate to 44100hz and extracted audio features. Kindly reply

@@venkatesanr9455 if you upsample, that would not hurt your signal ... But that would require more data storage to accommodate more samples, as @Valerio mentioned. Moreover, if your signal is bandpass signal of a carrier frequency (not a baseband), there is a limit of upsample as well, I mean you can't upsample your signal as much higher rate as you want

@@imamuddin8042 Thanks for your kind response, Sharif. I have recorded speech samples at sample rate of 8000 Hz. While processing this I had this doubt by upsampling the speech data towards 44.1k or 32kHz and extracting audio features would have any effects.

@@venkatesanr9455 Since you have recorded Speech Signal (that is 20Hz-20 Khz spanned) at sample rate of 8kHz , so already included aliasing effect of signals over 4khz... So I would recommend to put an anti aliasing low pass filter cut off at 4 khz in analog domain, before you sample your data. Or, you can initially use sampling rate 44.1 k , instead of thinking about upsampling it later

By far the most relevant one is wav

If you need general feedback, I suggest you to join The Sound of AI community (sign up link in the description). If you need more involved help, I do consulting.

Ok thank you

Indeed I'm a musician :) I play the piano.

@@ValerioVelardoTheSoundofAI That's wonderful