Motivation

Reported in #2986: batched inference through EncoderASR produces different transcripts than per-utterance inference, with divergences appearing early in the shorter (padded) utterances — not just in the tail.

Root cause: LengthsCapableSequential (and the lighter LengthsCapableChain) only forward lengths to layers whose forward() declares a parameter literally named lengths (speechbrain/utils/callchains.py::lengths_arg_exists). The HuggingFace integration lobes all name it wav_lens (forward(self, wav, wav_lens=None)), so in the standard encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential inference setup, the wav2vec2/HuBERT/WavLM module is silently called without lengths → make_padding_masks(src, wav_len=None) returns None → the HF model runs with attention_mask=None, and self-attention attends over the zero-padding, shifting every frame of the shorter items in the batch.

Note this is also a train/inference mismatch: training recipes typically call the lobe directly (self.modules.wav2vec2(wavs, wav_lens)), so the attention mask is applied during training but silently dropped at batched inference time.

Changes

• New speechbrain.utils.callchains.lengths_arg_name(func): returns the name of the relative-lengths argument (lengths or wav_lens), or None if there is none.
• LengthsCapableSequential.forward and LengthsCapableChain.__call__ now pass the lengths under the layer's actual parameter name.
• lengths_arg_exists is untouched, so the augmenter and enhancement call sites keep their exact behavior. The takes_lengths attributes are kept in sync (and now also report wav_lens-style layers).

Backward compatibility

• Layers with a lengths parameter: identical call as before.
• Layers with neither name: called without lengths, as before.
• Layers with a wav_lens parameter (the HF lobes): now actually receive the lengths — the intended fix. This deliberately changes batched-inference outputs for affected pipelines, aligning them with the mask-aware training-time behavior.

The issue reporter validated the equivalent fix end-to-end via a user-side wrapper: the WER delta between per-utterance and batched inference dropped to ~1e-3 over 300 samples (the residual is the convolutional front-end seeing the padding, inherent to padded batching). This PR makes that wrapper unnecessary.

Test Plan

• pytest tests/unittests/test_callchains.py tests/unittests/test_containers.py — 5 passed, including the new test_lengths_arg_name, test_lengths_capable_chain_wav_lens, and test_lengths_capable_sequential_lengths_dispatch (a wav_lens-named module actually receives the lengths through the container).
• pytest --doctest-modules speechbrain/nnet/containers.py speechbrain/utils/callchains.py tests/unittests/test_augment.py — 17 passed (augmenter unaffected).
• ruff check and ruff format --check clean on all changed files.

Fixes #2986.

Co-written by a human (@oliver0006) and Claude AI (Fable 5) working together.