fix(lid): align ECAPA inference with SpeechBrain

[beshkenadze]

Summary

• add sentence-level mean normalization before the ECAPA backbone
• align the ECAPA classifier path with SpeechBrain
• add regression coverage for the normalization and classifier ordering

Verification

• ./.venv/bin/python -m unittest mlx_audio.lid.tests.test_lid.TestEcapaTdnnModel.test_sentence_mean_normalize_centers_each_mel_bin mlx_audio.lid.tests.test_lid.TestEcapaTdnnModel.test_classifier_matches_speechbrain_order

Context

This is the follow-up fix after the shared ECAPA backbone landed in main. It keeps the MLX LID path aligned with the original SpeechBrain ECAPA inference behavior.

[Copilot]

Pull request overview

Aligns the MLX ECAPA-TDNN LID inference path with SpeechBrain’s behavior to reduce regression risk after introducing the shared ECAPA backbone.

Changes:

• Adds sentence-level mean normalization of mel features before the ECAPA backbone.
• Updates the ECAPA classifier path (LeakyReLU usage and ordering) to match SpeechBrain.
• Adds regression tests covering normalization behavior and classifier operation ordering.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
mlx_audio/lid/models/ecapa_tdnn/ecapa_tdnn.py	Applies sentence-mean normalization pre-backbone and aligns classifier activation/norm ordering with SpeechBrain.
mlx_audio/lid/tests/test_lid.py	Adds targeted regression tests for normalization centering and classifier ordering equivalence.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[beshkenadze]

Now we don't recognize accents the same way as in the original SpeechBrain 😁