fix(inference): make DiffWaveVocoder.forward and PIQAudioInterpreter.forward callable - #3077

What does this PR do?

Two Pretrained wrappers in speechbrain/inference/ call their own delegate
with an argument list the delegate cannot accept, so module(...) raises
TypeError before any inference happens. Both are one-line call-site
mismatches; I kept them in one PR because it is the same defect, but happy to
split if you prefer.

1. DiffWaveVocoder.forward (speechbrain/inference/vocoders.py)

def forward(self, spectrogram):
    return self.decode_batch(spectrogram)      # TypeError: missing a required argument: 'hop_len'

DiffWaveVocoder.decode_batch(self, mel, hop_len, mel_lens=None, ...) makes
hop_len required — it is passed to self.infer(scale=hop_len, ...) and to
mask_noise. The two sibling vocoders in the same file bind fine, which is
what makes this one stand out: HIFIGAN.decode_batch(self, spectrogram, mel_lens=None, hop_len=None) has hop_len optional, and
UnitHIFIGAN.forward forwards spk explicitly. The DiffWave wrapper looks
copied from HIFIGAN.forward without noticing that its decode_batch has a
second required parameter.

Fix: take hop_len in forward and pass it through.

2. PIQAudioInterpreter.forward (speechbrain/inference/interpretability.py)

def forward(self, wavs, wav_lens=None):
    return self.interpret_batch(wavs, wav_lens)   # TypeError: too many positional arguments

interpret_batch(self, wavs) takes no lengths, and the class's own
interpret_file already calls it as self.interpret_batch(batch). The
wav_lens parameter has therefore never reached anything — passing it (or
leaving it at its default) fails identically, because it is forwarded
positionally either way.

Fix: drop the argument that has no receiver, so forward matches
interpret_batch and interpret_file. I deliberately did not add
wav_lens support to interpret_batch instead: the method does no
length-based masking anywhere, so adding an ignored parameter would just move
the silent-drop one level down.

Neither class is referenced anywhere else in the repo (recipes/, tests/,
templates/, docs/ are all clean), and the docstring example in
interpretability.py uses interpret_file, so nothing that currently works
changes behaviour.

Fixes #<no issue — found by an arity sweep over self.<method>(...) call sites>

Verification

No model download or GPU needed. I replayed each forward -> delegate call
site against signatures parsed out of the real source with
inspect.Signature.bind, running the same harness over git show HEAD:<file>
and over the patched tree, and including the two sibling vocoders as a control:

===== BEFORE (git HEAD) =====
  FAIL  DiffWaveVocoder.forward(self, spectrogram) -> self.decode_batch(<spectrogram>)
          TypeError: missing a required argument: 'hop_len'
  OK    HIFIGAN.forward(self, spectrogram) -> self.decode_batch(<spectrogram>)
  OK    UnitHIFIGAN.forward(self, units, spk=None) -> self.decode_batch(<units>, spk=...)
  FAIL  PIQAudioInterpreter.forward(self, wavs, wav_lens=None) -> self.interpret_batch(<wavs>, <wav_lens>)
          TypeError: too many positional arguments

===== AFTER (patched working tree) =====
  OK    DiffWaveVocoder.forward(self, spectrogram, hop_len) -> self.decode_batch(<spectrogram>, <hop_len>)
  OK    HIFIGAN.forward(self, spectrogram) -> self.decode_batch(<spectrogram>)
  OK    UnitHIFIGAN.forward(self, units, spk=None) -> self.decode_batch(<units>, spk=...)
  OK    PIQAudioInterpreter.forward(self, wavs) -> self.interpret_batch(<wavs>)

Lint/format with the ruff version pinned in .pre-commit-config.yaml
(v0.12.4), run at the repo root:

ruff check speechbrain/inference/vocoders.py speechbrain/inference/interpretability.py   # All checks passed!
ruff format --check ...                                                                  # 2 files already formatted

🤖 Generated with Claude Code