Why there's big difference using 16k and 44.1k sample rate about python_speech_features HOT 5 CLOSED

For higher sample rates you'll need more filterbanks, otherwise each filterbank will be covering a large frequency range. If you do use more filterbanks, the higher ones will contain very little information since there is little speech info above 8khz. Sampling around 16khz should do the best.

from python_speech_features.

Actually, this comes together with another doubt I have. I am quite concerned about the pre-emphasis function - it seems that it is a basic subtraction of the original signal by the same signal slightly delayed (multiplied by a constant near 1). This delay is considerably different depending on the frame rate (delay = (62.5 uS for fr = 16k, 24.4us for fr = 44.1).

Do you guys have any suggestion of paper or any other technical reference about this specific approach of pre-emphasis?

from python_speech_features.

Preemphasis is used to 'flatten' the spectrum a little bit. For speech signals there is usually more energy in the low frequencies compared to the high frequencies. The Preemphasis filter is a highpass filter that evens out the energy a bit. It was used a lot in the past when euclidean distances were commonly used in asr systems. With gmms or neural nets preemphasis doesn't really matter, results will be the same whether you use it or not. It was included in the code because every other mfcc library I have seen includes it. You can safely ignore it for most purposes. Alternatively you can run some tests on a dev set with different preemph coefficients and see which, if any, works better.

from python_speech_features.

Thanks, James... Your response was really helpful

from python_speech_features.

No problem, glad I could help

from python_speech_features.