[magpietts] added an argument 'binarize_atten_prior' to trigger whether apply prior binarization.

[XuesongYang]

• added an arugment binarize_atten_prior to trigger whether apply prior binarization or not during training or evaluating. Previously, the training always applied binary prior. Now it is configurable.
• make prior past and future context window configurable during inference.
• hardcoded text_lens - 2 to decide if text sentence is finished or not.
• minor fixed typo..

[Copilot]

Pull Request Overview

This PR adds control over attention prior binarization and makes the prior context window and decay fully configurable for both training and inference, updates a hardcoded finish threshold, and fixes minor typos.

• Introduces binarize_attn_prior flag to toggle binarization of attention priors.
• Adds inference_prior_{future,past}_{context,decay} and inference_prior_current_value parameters for inference.
• Changes finish condition threshold from text_lens - 5 to text_lens - 2 and corrects typos.

Reviewed Changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
nemo/collections/tts/models/magpietts.py	Added binarize_attn_prior flag, configurable inference prior parameters, refactored attention prior construction, updated finish threshold, and typo fixes.
examples/tts/conf/magpietts/magpietts_multilingual_v1.yaml	Added binarize_attn_prior default to multilingual config.
examples/tts/conf/magpietts/magpietts_lhotse_dc_en.yaml	Added binarize_attn_prior default to lhotse DC English config.
examples/tts/conf/magpietts/magpietts_inference_multilingual_v1.yaml	Added binarize_attn_prior default to inference multilingual config.
examples/tts/conf/magpietts/magpietts_inference_en.yaml	Added binarize_attn_prior default to inference English config.
examples/tts/conf/magpietts/magpietts_en.yaml	Added binarize_attn_prior default to English config.
examples/tts/conf/magpietts/magpietts_dc_en.yaml	Added binarize_attn_prior default to DC English config.

Comments suppressed due to low confidence (5)

nemo/collections/tts/models/magpietts.py:286

• The newly added inference_prior_* parameters lack documentation. Please update the class or method docstring to explain their purpose and valid value ranges.

        # Inference prior configuration

nemo/collections/tts/models/magpietts.py:1222

• Since behavior now changes based on binarize_attn_prior, add unit tests covering both True and False cases to ensure the logic branches produce the expected prior matrices.

                if self.binarize_attn_prior:

nemo/collections/tts/models/magpietts.py:289

• The attribute prior_future_decay is not defined before use. Consider obtaining it from cfg or defining a default value earlier to avoid an AttributeError.

        self.inference_prior_future_decay = self.cfg.get('inference_prior_future_decay', self.prior_future_decay)

nemo/collections/tts/models/magpietts.py:290

• The attribute prior_past_decay is not defined before use. Consider obtaining it from cfg or defining a default value earlier to avoid an AttributeError.

        self.inference_prior_past_decay = self.cfg.get('inference_prior_past_decay', self.prior_past_decay)

nemo/collections/tts/models/magpietts.py:1222

• When binarize_attn_prior is False, aligner_attn_hard is never assigned but is likely used later. Add an else branch or default assignment to prevent an UnboundLocalError.

                if self.binarize_attn_prior:

[Copilot]

[nitpick] This uses a magic number (-2) to detect sentence completion. Consider extracting it into a named constant or making it configurable for clarity and easier tuning.

Suggested change

	if text_time_step_attended[bidx] >= text_lens[bidx] - 2 or bidx in end_indices:
	if text_time_step_attended[bidx] >= text_lens[bidx] - self.EOS_WINDOW_OFFSET or bidx in end_indices: