[michalharakal]

MelSpectrogram computed each frame's power spectrum with a direct DFT over precomputed cos/sin tables: O(n²/2) per frame and 640 KB of tables for Whisper's nFft = 400. This replaces it with an FFT.

What changes

• PowerSpectrum (internal): a mixed-radix FFT for any frame size.
  • Radix 4, 2 and 5 butterflies are written out in closed form; radix 3 and larger factors use the plain sum. Whisper's 400-point frame is 4·4·5·5, so no power-of-two size is needed.
  • Twiddles are computed in double precision. The buffers are allocated per logMel call, so MelSpectrogram is still safe to share across threads.
• Each mel filter is summed only over its non-zero band. This is exact, because the skipped terms are 0.

Numbers

	Before	After
JVM, one 4 s Whisper window	38 ms	5.8 ms
ART on an arm32 Android device (log-mel inside a Whisper tiny transcription, same transcript)	≈ 0.88 s	≈ 0.2 s
whisper.log_mel_spectrogram golden (MelSpectrogramTest)	cosine 0.99999998, max abs diff 1.3e-5	cosine 0.99999999, max abs diff 1.5e-5

Tests
PowerSpectrumTest in commonTest, green on jvm and linuxX64:

• the power spectrum against a double-precision direct DFT for 400, 512, every radix path, primes and 7·7
• a pure tone lands in its bin
• the factorization

The existing mel tests pass unchanged.