[TTS][Magpietts] Unify Longform and Standard Inference logic

[subhankar-ghosh]

Important

The Update branch button must only be pressed in very rare occassions.
An outdated branch is never blocking the merge of a PR.
Please reach out to the automation team before pressing that button.

What does this PR do ?

This pull request refactors and unifies the text chunking and inference logic for TTS (Text-to-Speech) in the MagpieTTS pipeline. The main change is the replacement of the previous "longform" inference logic with a new, language-aware, unified chunked inference path. This affects dataset preparation, model state management, argument parsing, and the inference runner, making the codebase simpler and more robust for both short and long texts.

Key changes:

Unified Inference and Text Chunking

• Replaced the old longform inference logic with a unified, automatic text chunking approach that determines chunking based on per-sample language thresholds. Short texts are processed as single chunks, while long texts are split into sentences automatically. (examples/tts/magpietts_inference.py, nemo/collections/tts/data/text_to_speech_dataset.py, nemo/collections/tts/models/magpietts.py) [1] [2] [3]
• Removed all command-line arguments related to explicit longform control (--longform_mode, --longform_word_threshold, etc.), simplifying the inference interface. (examples/tts/magpietts_inference.py) [1] [2] [3]

Dataset and Collation Refactor

• Introduced ChunkedTTSInferenceDataset (replacing LongFormTTSInferenceDataset) with per-sample, language-aware chunking and tokenizer selection. The dataset now automatically decides chunking strategy based on language and text length. (nemo/collections/tts/data/text_to_speech_dataset.py) [1] [2] [3]
• Updated the dataset's collate_fn to handle variable-length chunked batches, padding as needed, and to generalize beyond the previous longform-specific logic. (nemo/collections/tts/data/text_to_speech_dataset.py) [1] [2]

Model and State Naming Consistency

• Renamed all "longform" classes and configs to "chunked" (e.g., LongformDecoderState → ChunkedDecoderState, LongformConfig → ChunkedInferenceConfig) throughout the model code for clarity and consistency with the new unified approach. (nemo/collections/tts/models/magpietts.py) [1] [2] [3] [4]
• Removed the _needs_longform_inference method and all language threshold logic from the model, as chunking is now handled in a unified, language-aware way. (nemo/collections/tts/models/magpietts.py)

Utility and Import Updates

• Added and updated utility imports for chunked inference and tokenizer selection to support the new pipeline. (nemo/collections/tts/models/magpietts.py, nemo/collections/tts/data/text_to_speech_dataset.py) [1] [2]

These changes make the TTS inference pipeline easier to use and maintain, while improving support for multilingual and variable-length text inputs.

Collection: TTS

Changelog

• Add specific line by line info of high level changes in this PR.

Usage

• You can potentially add a usage example below

# Add a code snippet demonstrating how to use this 

GitHub Actions CI

The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.

The GitHub Actions CI will run automatically when the "Run CICD" label is added to the PR.
To re-run CI remove and add the label again.
To run CI on an untrusted fork, a NeMo user with write access must first click "Approve and run".

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

• Related to # (issue)

[Copilot]

Pull request overview

Refactors MagpieTTS inference to use a single “chunked” inference path for both short and long texts, with dataset-driven automatic sentence chunking based on per-sample language thresholds.

Changes:

• Introduces language-aware thresholding + unified chunk_text_for_inference() chunking utility (replacing prior longform detection logic).
• Replaces LongFormTTSInferenceDataset with ChunkedTTSInferenceDataset and updates the inference runner to always use the unified multi/single-chunk loop.
• Updates CLI/example script to remove explicit longform args and align with the unified inference flow.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 8 comments.

Show a summary per file

File	Description
nemo/collections/tts/parts/utils/tts_dataset_utils.py	Adds language-aware sentence splitting, thresholds, tokenizer mapping, and unified chunking helper.
nemo/collections/tts/data/text_to_speech_dataset.py	Replaces longform inference dataset with unified chunked inference dataset + mixed-chunk collation.
nemo/collections/tts/modules/magpietts_inference/inference.py	Removes standard/longform branching; always runs unified chunk loop via generate_speech().
nemo/collections/tts/models/magpietts.py	Renames longform state/config to chunked equivalents; updates do_tts() to the unified chunked generation path.
examples/tts/magpietts_inference.py	Removes longform CLI controls and updates messaging for unified chunking behavior.
tests/collections/tts/parts/utils/test_tts_dataset_utils.py	Adds unit tests for new thresholds and unified chunking helper.

---

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

[blisc]

• Can you resolve conflicts?
• Can you add a unit test for magpietts.generate_speech that contains a batch of short and long texts?

[blisc]

The github UI still says that there are conflicts

[blisc]

Please fix the linting errors

[blisc]

If this is the same checkpoint as tests/functional_tests/L2_TTS_InferEvaluate_Magpietts_SeenSpeakers.sh, we can remove this test. This test was only kept to test a different checkpoint

[blisc]

We can remove this test now since it's identical to L2_TTS_InferEvaluate_Magpietts_SeenSpeakers.sh

[subhankar-ghosh]

Fixed the necessary linting issues, the current linting failures are due to CI issues.

[github-actions]

[🤖]: Hi @subhankar-ghosh 👋,

We wanted to let you know that a CICD pipeline for this PR just finished successfully.

So it might be time to merge this PR or get some approvals.