New ML yaml + changes to allow for Spectral Codec training with text context

[blisc]

No description provided.

[Copilot]

Pull Request Overview

This PR introduces changes to support Spectral Codec training with text context in the Magpie-TTS model. The changes include modifications to codec model loading to disable loss modules during inference and variable reorganization to handle different codebook configurations.

• Codec model loading enhancement to disable SCL loss during inference for memory optimization
• Variable restructuring to distinguish between data and model codebook configurations
• New multilingual configuration file for Magpie-TTS with text conditioning support

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
nemo/collections/tts/models/magpietts.py	Modified codec loading logic and reorganized codebook variables to support different training scenarios
examples/tts/conf/magpietts/magpietts_multilingual_v2_lhotse.yaml	Added new configuration file for multilingual Magpie-TTS with text conditioning

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[rlangman]

Looks good to me.

[rlangman]

I am not sure if there is a clean way to handle reflect/replicate padding when input is short. We could have each codec architecture define a minimum length it can handle, and then have pad_audio zero pad to at least that length: https://github.com/blisc/NeMo/blob/magpietts_2503/nemo/collections/tts/models/audio_codec.py#L453

[blisc]

I agree, we should do that.