We make sure the CELT PLC we do in the transition uses the same bandwidth
as the previous (CELT) packet and not the new bandwidth

Andrew Allen authored 7 years ago and Jean-Marc Valin committed 7 years ago

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>

A second-order DC rejection filter is uselsss unless we have complex
poles. However, complex poles means we have to compute the filter as a
single pass (rather than two casdaded first-order filters), which has
numerical issues that would require a higher complexity to solve.
So rather than waste cycles with a second-order filter (with a longer
impulse response), we just go with a first-order filter.

Fixes a "regression" from 1.2 to 1.3-beta that was especially noticeable
on a Korean speech sample. (it wasn't a real regression since 1.2 only
worked because it was using CELT)

That corresponds to the fundamental for the shortest pitch period allowed

Andrew Allen authored 7 years ago and Jean-Marc Valin committed 7 years ago

Change-Id: I3db285875b6b5548decc436be00096b97be1be3c
Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>

Andrew Allen authored 7 years ago and Jean-Marc Valin committed 7 years ago

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>

Andrew Allen authored 7 years ago and Jean-Marc Valin committed 7 years ago

Modified by Jean-Marc Valin

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>

Andrew Allen authored 7 years ago and Jean-Marc Valin committed 7 years ago

Ensure matrix is properly exported from
projection encoder.

Change-Id: I5ede77e4f4d0dc82074e2230780777af542a416e
Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>

Larger network, using 8-bit weights

Andrew Allen authored 7 years ago and Jean-Marc Valin committed 7 years ago

OpusProjection* classes
MixingMatrix class
Projection tests

Change-Id: I98644466abf4ffd36e48bdecad1204d69e1539b9
Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>

The masking term was previously completely broken (even in 1.1). The
bandwidth detection relies mostly on the noise floor and can only use
masking to cut one extra band. The 12-24 kHz energy is now normalized properly
but uses a higher noise floor due to the leakage in the resampler. Bandwidth
detection is still mostly useless at identifying SWB speech (it usually says
FB).

Based on two dense layers with a GRU layer in the middle

SILK was being allocated too few bits for stereo hybrid, often resulting
in forced narrowing of the width. We now allocate more bits to SILK
and reduce the threshold for narrowing. This improves quality enough that
the bitrate threshold for switching to SILK can be increased to 44 kb/s.

Matt Brubeck authored 7 years ago and Ralph Giles committed 7 years ago

Signed-off-by: Ralph Giles <giles@thaumas.net>

Andrew Larkin authored 7 years ago and Felicia Lim committed 7 years ago

Signed-off-by: Felicia Lim <flim@google.com>

The new detector for SWB vs FB is unreliable due to the hack
that was required to get 24 kHz analysis working. We're now err
on the side of FB just to make sure.

Linfeng Zhang authored 8 years ago and Jean-Marc Valin committed 7 years ago

Should call celt_inner_prod().

This change is bit exact as original, except for x86 floating-point.
In x86 floating-point, it calls celt_inner_prod_sse() which may have
different output with the change of floating-point operations' orders.

Change-Id: Ia2381e2e198a84296ac28305183f15be842b3454

Signed-off-by: Jean-Marc Valin <jmvalin@jmvalin.ca>

This was sometimes causing the music probability to be stuck at
zero at the beginning of files

We boost bands that either cause leakage or are filled with leakage

Now scaling the energy to the same value is would be in floating-point

This makes it possible to run the testvectors with asan enabled.

The constant was copied from the analysis code which uses +/-1 as float
scaling, but surround_analysis() uses +/-32767.