Fix normalization of impulse response in ImpulsePerturbation

[anteju]

What does this PR do?

Current impulse response (IR) in ImpulsePerturbation is not correct.
This PR fixes the issue.

Check below for

1. Audio examples
2. IR examples
3. Scripts to reproduce the issue

Collection: ASR

Changelog

• Fixed normalization to correctly normalize mean and max
• Added a new flag to make IR normalization optional
• Default to not using impulse normalization
  1. This doesn't negatively impact the current behavior, since currently IR is incorrectly normalized
  2. Normalization is in general not necessary for IRs, especially since output data is normalized to max amplitude 1
  3. If necessary, normalization can be turned on to remove the mean from the IR.
• fixed convolution and impulse shifting
  • shift_impulse=False used same convolution, which centered the output with respect to full. This resulted in an output which was not aligned with the RIR. This is now done by computing the full convolution and trimming the output. This is a bugfix.
  • shift_impulse=True used trimmed RIR. This is now done by convolving with the full RIR and shifting the output. This is a minor change.

Detailed description

Current impulse response (IR) in ImpulsePerturbation is not correct.
More specifically, it introduces a DC offset in the normalized IR, resulting in an incorrect signal.
It's obvious from listening to the output audio that it doesn't correspond to a realistic IR.

Example scripts

• rir_normalization_issue_example.zip

To run, unpack the archive and run

python example.py

RIR example

This figure shows an original IR, the normalized IR with the current normalization and the normalized IR with the new normalization (in this PR).

The current normalization introduces a large DC offset.
The new normalization makes the RIR zero mean, and scales the max amplitude to 1.

Audio examples

Output audio with the current normalization

• output_cur.wav.zip

Output audio with the new normalization

• output_new.wav.zip

Test with different setups

Scripts

rir_normalization_issue_example_with_shift.zip

In this example, the RIR introduces a 0.5 second delay.
This was artificially constructed to test the setup with shift_impulse.
We expect the following

• shift_impulse=False: The output should be delayed by 0.5s due to delay in the RIR
• shift_impulse=True: The output should not include the 0.5s delay

Input signal

Output signals

	shift_impulse=False	shift_impulse=True
normalize_impulse=False
normalize_impulse=True

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

• Related to Impulse perturbation normalization typo #6429