[ktz03]

log_mel_spectrogram accepts NumPy waveforms, but a reversed or negatively strided view currently fails in torch.from_numpy before computing the spectrogram. An equivalent contiguous array works:

import numpy as np
import whisper

time = np.arange(3200, dtype=np.float32) / 16000
audio = np.sin(2 * np.pi * 440 * time).astype(np.float32)[::-1]
whisper.log_mel_spectrogram(audio)

Copy the waveform only when it has a negative stride before converting it to a tensor. Contiguous and positive-stride arrays retain the existing conversion path. The added regressions compare positive and negative strides with the same waveform values for both 80 and 128 Mel filters, and check the input, output dtype, and CPU device.

Validation on Windows, Python 3.12.14, PyTorch 2.14.1+cpu, NumPy 2.5.3:

• The new regressions fail 4 cases on unmodified main and pass all 8 with the fix; the 4 positive-stride controls pass on both.
• pytest tests/test_audio.py -q: 9 passed, including the bundled audio file decoded with FFmpeg.
• The CPU test selection from .github/workflows/test.yml: 27 passed, 20 deselected, with the expected 2 CPU FP16 fallback warnings. This includes real tiny and tiny.en transcription with word timestamps using the official weights, verified against their SHA256 hashes.
• The repository's pinned pre-commit run --all-files: passed.

GPU execution, other model sizes, the complete Python/PyTorch CI matrix, and performance were not tested locally.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-09T13:54:56.335961Z	af031cc	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.