fix(whisper): match reference mel preprocessing

[WenJing95]

Background

Nonzero audio edges produced different Whisper mel features because STFT used zero padding. The discarded final frame could also change the log floor.

Exit Criteria

Match reference features at chunk boundaries.

Implementation

Reflect both ends of the waveform and remove the last frame before normalization.

Change categories

• Model or runtime behavior

Validation

Commands and Results

Native reference goldens passed at 80/128 mels; source-quality: 164 passed.

g++ -std=c++17 -O2 -Wall -Wextra -Wpedantic -I. families/whisper/tests/cpp/test_whisper_mel_spectrogram.cpp families/whisper/runtime/whisper_mel_spectrogram.cpp -o /tmp/trtmc-four-whisper-test
/tmp/trtmc-four-whisper-test
python -m tools.community_ci source-quality --base 714f1fc0d567213a7b79a488e9dcfdf405279c6a

Hardware, Environment, and Revisions

8a95b0e: Ubuntu CPU, GCC 13.3; Transformers 5.2.0 reference.

Not Run / Remaining Gaps

TensorRT inference and full CMake suite: TensorRT environment unavailable.

Contributor Self-Review

• I have completed a self-review of this change.

Notes For Future Readers

Reflect after padding/truncating to 30 seconds so short clips keep their zero-padded tail.

Risk level

• Low

Changes are confined to Whisper's center padding and log normalization.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

_{Comment @coderabbitai help to get the list of available commands.}