Add support for scipy polyphase resampling

Description

Scipy 0.18 added a polyphase resampler. We could add support for this to core.resample, eg:

y_res = librosa.resample(y, sr_old, sr_new, res_type='polyphase')

Potential complications

1. This would up the version requirement for scipy from 0.13 to 0.18. We could potentially make this soft.
2. We’d probably need to expose some new parameters for specifying the window.
3. We’d need a good way to determine the up and down parameters. If the sampling rates are integer-valued, we can figure this out from the gcd. However, the current API does not require integer sampling rates, and I’d like to keep that as flexible as possible. Perhaps we could only require int-valued rates for this particular mode, and throw a warning / exception if fractional rates are provided.

[dpwe]

FWIW, Matlab’s resample also requires integer ratios. I wrote code to find ratios to get arbitrarily close to a resample ratio expressed as a float (I was resampling audio by factors very close to 1 to get time alignment between simultaneous recordings with unsynchronized sample clocks). See the first block (calculating the up/down factors P and Q) in:

https://labrosa.ee.columbia.edu/projects/skewview/rewrite_aligned_audio.m.html

DAn.

On Wed, May 17, 2017 at 10:47 AM, Brian McFee [e-mail] wrote:

Description

Scipy 0.18 added a polyphase resampler https://docs.scipy.org/doc/scipy-0.18.1/reference/generated/scipy.signal.resample_poly.html. We could add support for this to core.resample, eg:

y_res = librosa.resample(y, sr_old, sr_new, res_type=‘polyphase’)

Potential complications

1. This would up the version requirement for scipy from 0.13 to 0.18. We could potentially make this soft.
2. We’d probably need to expose some new parameters for specifying the window.
3. We’d need a good way to determine the up and down parameters. If the sampling rates are integer-valued, we can figure this out from the gcd. However, the current API does not require integer sampling rates, and I’d like to keep that as flexible as possible. Perhaps we could only require int-valued rates for this particular mode, and throw a warning / exception if fractional rates are provided.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/librosa/librosa/issues/576, or mute the thread https://github.com/notifications/unsubscribe-auth/AAhs0TKP0FF-np7ChSvP741388LK0xwKks5r6wiEgaJpZM4Nd-eq .

[bmcfee]

note: i wound up adding this in #859 primarily to give folks the option of having fast resampling in (i)cqt if they want it.n I opted to only implement integer rates, which are automatically reduced by GCD to minimize overhead.