fix(asr): pad random_segment with an integer sample count

[udsy19]

What does this PR do ?

Fixes RandomSegmentPerturbation (random_segment augmentation) crashing with
TypeError: pad_width must be of integral type. whenever it has to pad a short utterance, and
corrects the SSL configuration doc that documents the padding behaviour but omits the flag that
enables it.

Collection: ASR

Changelog

• nemo/collections/asr/parts/preprocessing/perturb.py: compute pad_size as an integer number
  of samples with math.ceil, so AudioSegment.pad -> numpy.pad receives an integral
  pad_width. Previously duration_sec * sample_rate - num_samples was a float whenever
  duration_sec was a float — including the constructor default 32.0 — and every short
  utterance raised TypeError. math.ceil rather than round/truncation is deliberate: the
  next statement is subsegment(0.0, duration_sec), which raises ValueError if the segment
  ends up shorter than duration_sec, so the padded length must not round down. subsegment
  then trims to exactly round(duration_sec * sample_rate) samples, matching the non-padding
  branch.
• docs/source/asr/ssl/configs.rst: add pad_to_duration: true to the random_segment snippet.
  The surrounding text says "samples below the provided segment length will be padded", but the
  snippet omitted the flag and it defaults to False, so the documented recipe silently did not
  pad — which defeats its stated purpose of keeping durations uniform inside an SSL batch.
• tests/collections/asr/test_preprocessing_segment.py: add TestRandomSegmentPerturbation
  covering int and float duration_sec, the zero-fill placement, sample rates where
  duration_sec * sample_rate is not a whole number (22050 Hz), the pad_to_duration=False
  default, and the non-padding branch. There was previously no test for this perturbation.

Usage

import numpy as np
from nemo.collections.asr.parts.preprocessing.perturb import RandomSegmentPerturbation
from nemo.collections.asr.parts.preprocessing.segment import AudioSegment

seg = AudioSegment(np.zeros(16000 * 5, dtype=np.float32), 16000)      # 5 s
RandomSegmentPerturbation(duration_sec=32.0, pad_to_duration=True).perturb(seg)
len(seg.samples)   # before: TypeError    after: 512000

Equivalently, from a training config:

model:
  train_ds:
    augmentor:
      random_segment:
        prob: 1.0
        duration_sec: 16.0
        pad_to_duration: true

GitHub Actions CI

Checks run locally (macOS, CPU, Python 3.10.18, torch 2.12.0, numpy 2.2.6):

• pre-commit run --from-ref main --to-ref HEAD — all hooks pass (isort, black 24.10.0,
  check-case-conflict, detect-private-key, check-added-large-files).
• pytest tests/collections/asr/test_preprocessing_segment.py -m "not pleasefixme" — 73 passed.
• pytest tests/collections/asr -m "not pleasefixme" — 1402 passed, 402 skipped, 26 failed,
  12 errors. All 38 failures/errors are pre-existing on main and environmental on a CPU-only
  macOS box (*_gpu parametrisations, ONNX/TorchScript export, model-download errors); none is
  in preprocessing or perturbation code. Verified by re-running the 7 affected files with
  perturb.py reverted to main: the failure sets are byte-identical, so this change
  introduces no new failures.
• Regression check: with perturb.py reverted to main and the new tests in place, 6 of the 10
  new cases fail (TypeError: pad_width must be of integral type.); the 4 that pass are the
  int-duration_sec, pad_to_duration=False and long-audio controls. All 10 pass with the fix.
• git diff --check — clean.
• Not run: the Sphinx docs build. The doc change is one line inside an existing
  .. code-block:: yaml and adds no directive, role or reference; the snippet was parsed with
  yaml.safe_load to confirm it is still valid.

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)
  • No. math is stdlib; no new dependency and no optional import.

PR Type:

• New Feature
• Bugfix
• Documentation

Additional Information

• Fixes random_segment augmentation raises TypeError: pad_width must be of integral type whenever it has to pad #16104
• The documentation change is included here rather than split out because it is the same
  user-visible failure: on main the documented random_segment recipe does not pad, and the
  doc line would be actively misleading — it would walk readers into the TypeError — if landed
  without the code fix.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.