Add Kannada (kn-IN) G2P support for TTS by jasro23 · Pull Request #15582 · NVIDIA-NeMo/NeMo

jasro23 · 2026-04-03T19:13:05Z

Add KannadaG2p class with hybrid dictionary + rule-based IPA conversion
Add Kannada grapheme and IPA character sets to ipa_lexicon.py
Add kn-IN locale support with punctuation handling
Include lexicon with 4K Kannada words
Add test script with assertions for validation

The G2P module handles:

All Kannada vowels, consonants, matras (dependent vowels)
Virama (halant), anusvara, visarga
Anusvara place assimilation based on following consonant

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An outdated branch is never blocking the merge of a PR.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 ?

Add Kannada (kn-IN) G2P support for TTS

Collection: [TTS]

Changelog

Add KannadaG2p class in nemo/collections/tts/g2p/models/kn_in_ipa.py for Kannada grapheme-to-phoneme conversion
Add Kannada grapheme and IPA character sets to ipa_lexicon.py
Add kn-IN to supported locales with Kannada-specific punctuation
Add pronunciation dictionary with 4264 Kannada words (kn_IN_nv260318.dict)
Add Kannada G2P unit tests to test_tts_tokenizers.py

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The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.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

- Add KannadaG2p class with hybrid dictionary + rule-based IPA conversion - Add Kannada grapheme and IPA character sets to ipa_lexicon.py - Add kn-IN locale support with punctuation handling - Include lexicon with 4264 Kannada words - Add test script with assertions for validation The G2P module handles: - All Kannada vowels, consonants, matras (dependent vowels) - Virama (halant), anusvara, visarga - Anusvara place assimilation based on following consonant Signed-off-by: Jason Roche <jas.tech23@gmail.com>

Signed-off-by: jasro23 <jasro23@users.noreply.github.com>

annagirimokshith · 2026-04-05T06:31:42Z

@jasro23 Can you also add support for Telugu language

jasro23 · 2026-04-09T15:50:09Z

@jasro23 Can you also add support for Telugu language

@annagirimokshith . I am not too familiar with Telugu.

+# limitations under the License.
+
+import pathlib
+import re


+from nemo.collections.common.tokenizers.text_to_speech.ipa_lexicon import (
+    GRAPHEME_CHARACTER_SETS,
+    get_grapheme_character_set,
+    get_ipa_punctuation_list,
+)


Copilot

Pull request overview

Adds Kannada (kn-IN) grapheme-to-phoneme (G2P) support for NeMo TTS, including a new Kannada IPA G2P implementation, locale character sets/punctuation, a pronunciation dictionary, and unit tests.

Changes:

Introduce KannadaG2p with hybrid dictionary + rule-based IPA conversion.
Add kn-IN grapheme and IPA character sets plus locale punctuation handling.
Add a Kannada pronunciation lexicon and basic unit tests validating G2P outputs.

Reviewed changes

Copilot reviewed 3 out of 4 changed files in this pull request and generated 6 comments.

File	Description
`nemo/collections/tts/g2p/models/kn_in_ipa.py`	New Kannada G2P implementation (dictionary + rule-based).
`nemo/collections/common/tokenizers/text_to_speech/ipa_lexicon.py`	Adds `kn-IN` locale support, including grapheme/IPA sets and punctuation.
`scripts/tts_dataset_files/kn_IN/kn_IN_nv260318.dict`	New Kannada pronunciation dictionary (~4.3K entries).
`tests/collections/common/tokenizers/text_to_speech/test_tts_tokenizers.py`	Adds unit tests for Kannada G2P behavior.

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

Copilot · 2026-04-20T18:33:22Z

+            # Handle digits (pass through or convert)
+            if char.isdigit():
+                phonemes.append(char)
+                i += 1
+                continue
+
+            # Handle Kannada digits
+            kannada_digits = '೦೧೨೩೪೫೬೭೮೯'
+            if char in kannada_digits:
+                # Convert to Arabic numeral
+                phonemes.append(str(kannada_digits.index(char)))
+                i += 1
+                continue


Kannada digits (೦-೯) will never reach the Kannada-digit conversion block because str.isdigit() is true for Kannada digits, so the earlier if char.isdigit(): phonemes.append(char) branch consumes them. This causes Kannada digits to be returned unchanged instead of being mapped to Arabic numerals as intended. Consider checking for Kannada digits before isdigit(), or restricting the isdigit() branch to ASCII digits only (e.g., char.isascii() and char.isdigit()).

Copilot · 2026-04-20T18:33:22Z

+            phoneme_dict = (
+                self._parse_phoneme_dict(phoneme_dict, phoneme_prefix)
+                if isinstance(phoneme_dict, (str, pathlib.Path))
+                else phoneme_dict
+            )


When phoneme_dict is passed as a Python dict, entries like {word: ["namaskaːɾa"]} are kept as whole-string tokens, but the rule-based path emits per-character IPA tokens. This makes outputs inconsistent across dictionary vs OOV words and can break downstream tokenizers that expect single-symbol IPA tokens (e.g., 'ː' separate from 'a'). Consider normalizing dict-provided pronunciations by splitting each pronunciation string into a list of IPA symbols/characters (and applying phoneme_prefix) to match _parse_phoneme_dict behavior.

Suggested change

phoneme_dict = (

self._parse_phoneme_dict(phoneme_dict, phoneme_prefix)

if isinstance(phoneme_dict, (str, pathlib.Path))

else phoneme_dict

)

if isinstance(phoneme_dict, (str, pathlib.Path)):

phoneme_dict = self._parse_phoneme_dict(phoneme_dict, phoneme_prefix)

else:

normalized_phoneme_dict = {}

for word, prons in phoneme_dict.items():

normalized_prons = []

for pron in prons:

if isinstance(pron, str):

normalized_prons.extend([phoneme_prefix + symbol for symbol in pron])

else:

normalized_prons.extend(pron)

normalized_phoneme_dict[word] = normalized_prons

phoneme_dict = normalized_phoneme_dict

Copilot · 2026-04-20T18:33:22Z

+        Args:
+            phoneme_dict: Path to Kannada pronunciation dictionary file or a dict object.
+                Format: word<whitespace>phonemes (space-separated IPA symbols)
+            phoneme_prefix: Prefix to prepend to phoneme symbols to distinguish from graphemes.
+                Default is "" (no prefix).
+            ascii_letter_prefix: Prefix to prepend to ASCII letters for code-mixed text.
+                Default is "" (no prefix).
+            ascii_letter_case: Case for ASCII letters: "upper", "lower", or "mixed".
+                Default is "lower".
+            word_tokenize_func: Custom function for tokenizing text into words.
+                Should return List[Tuple[Union[str, List[str]], bool]].
+            apply_to_oov_word: Custom function to apply to out-of-vocabulary words.
+                If None, rule-based G2P is used.
+            mapping_file: Optional path to character mapping file.


The docstring states the dictionary format uses “space-separated IPA symbols”, but _parse_phoneme_dict() (and the provided .dict file) treat pronunciations as a single IPA string without spaces and then split into characters. Please align the documentation with the actual expected format to avoid confusing users providing custom dictionaries.

Copilot · 2026-04-20T18:33:23Z

+import re
+import unicodedata
+from collections import defaultdict
+from typing import Dict, List, Optional, Union
+
+from nemo.collections.common.tokenizers.text_to_speech.ipa_lexicon import (
+    GRAPHEME_CHARACTER_SETS,


re and GRAPHEME_CHARACTER_SETS are imported but not used in this module. Removing unused imports will avoid lint failures and keep the file clean.

Suggested change

import re

import unicodedata

from collections import defaultdict

from typing import Dict, List, Optional, Union

from nemo.collections.common.tokenizers.text_to_speech.ipa_lexicon import (

GRAPHEME_CHARACTER_SETS,

import unicodedata

from collections import defaultdict

from typing import Dict, List, Optional, Union

from nemo.collections.common.tokenizers.text_to_speech.ipa_lexicon import (

Copilot · 2026-04-20T18:33:23Z

+        '್',  # Virama (Halant)
+        'ಽ',  # Avagraha


The kn-IN grapheme character set omits some characters that the new Kannada G2P code references/handles (e.g., nukta '಼') and common script punctuation (e.g., danda '।', double danda '॥'). If these are expected in input text, consider adding them to GRAPHEME_CHARACTER_SETS["kn-IN"] (similar to hi-IN) so downstream grapheme-based processing/tokenizers can recognize them.

Suggested change

'್', # Virama (Halant)

'ಽ', # Avagraha

'್', # Virama (Halant)

'಼', # Nukta

'ಽ', # Avagraha

# Script punctuation

'।', # Danda

'॥', # Double danda

Copilot · 2026-04-20T18:33:23Z

+    @pytest.mark.run_only_on('CPU')
+    @pytest.mark.unit
+    def test_kannada_g2p_rule_based(self):
+        """Test Kannada G2P rule-based conversion."""
+        g2p = KannadaG2p()
+
+        # Test basic vowels
+        assert "".join(g2p("ಅ")) == "a"
+        assert "".join(g2p("ಆ")) == "aː"
+
+        # Test consonant with inherent vowel
+        assert "".join(g2p("ಕ")) == "ka"
+
+        # Test consonant with virama (no vowel)
+        assert "".join(g2p("ಕ್")) == "k"
+
+        # Test consonant with matra
+        assert "".join(g2p("ಕಾ")) == "kaː"
+        assert "".join(g2p("ಕಿ")) == "ki"
+
+        # Test word
+        assert "".join(g2p("ಕನ್ನಡ")) == "kannaɖa"
+
+    @pytest.mark.run_only_on('CPU')
+    @pytest.mark.unit
+    def test_kannada_g2p_dictionary_based(self):
+        """Test Kannada G2P with dictionary lookup."""
+        g2p = KannadaG2p(phoneme_dict=self.PHONEME_DICT_KN)
+
+        # Test dictionary lookup
+        assert "".join(g2p("ನಮಸ್ಕಾರ")) == "namaskaːɾa"
+        assert "".join(g2p("ಕನ್ನಡ")) == "kannaɖa"
+        assert "".join(g2p("ಬೆಂಗಳೂರು")) == "beŋgaɭuːɾu"
+
+    @pytest.mark.run_only_on('CPU')
+    @pytest.mark.unit
+    def test_kannada_g2p_special_characters(self):
+        """Test Kannada G2P with special characters."""
+        g2p = KannadaG2p()
+
+        # Test anusvara (nasal)
+        result = "".join(g2p("ಅಂಕ"))
+        assert "ŋ" in result  # Anusvara before velar should be ŋ
+
+        # Test retroflex consonants
+        assert "ɖ" in "".join(g2p("ಡ"))
+        assert "ɳ" in "".join(g2p("ಣ"))
+
+        # Test vocalic R
+        assert "ɾɯ" in "".join(g2p("ಕೃ"))
+
+    @pytest.mark.run_only_on('CPU')
+    @pytest.mark.unit
+    def test_kannada_g2p_affricates(self):
+        """Test Kannada G2P affricate handling."""
+        g2p = KannadaG2p()
+
+        # Test palatal affricates
+        result = "".join(g2p("ಚ"))
+        assert "tʃ" in result or ("t" in result and "ʃ" in result)
+
+        result = "".join(g2p("ಜ"))
+        assert "dʒ" in result or ("d" in result and "ʒ" in result)


The new Kannada G2P tests don’t cover Kannada digit handling (೦-೯) or punctuation passthrough, even though the implementation includes explicit logic for both. Adding assertions for these cases would prevent regressions (and would have caught the current Kannada-digit bug).

github-actions Bot added TTS common labels Apr 3, 2026

jasro23 force-pushed the kannada_g2p branch from 8a831ca to 23b2422 Compare April 3, 2026 19:47

jasro23 force-pushed the kannada_g2p branch from 2cb28e8 to 8ddb3c2 Compare April 3, 2026 19:59

Apply isort and black reformatting

aa14c48

Signed-off-by: jasro23 <jasro23@users.noreply.github.com>

github-actions Bot added the community-request label Apr 5, 2026

chtruong814 added the needs-follow-up Issue needs follow-up label Apr 11, 2026

XuesongYang self-requested a review April 20, 2026 17:41

XuesongYang added the Run CICD label Apr 20, 2026

XuesongYang requested a review from Copilot April 20, 2026 18:27

Copilot started reviewing on behalf of XuesongYang April 20, 2026 18:28 View session

github-advanced-security AI found potential problems Apr 20, 2026

View reviewed changes

Copilot AI reviewed Apr 20, 2026

View reviewed changes

svcnvidia-nemo-ci added waiting-on-maintainers Waiting on maintainers to respond and removed needs-follow-up Issue needs follow-up labels Apr 21, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add Kannada (kn-IN) G2P support for TTS#15582

Add Kannada (kn-IN) G2P support for TTS#15582
jasro23 wants to merge 2 commits intoNVIDIA-NeMo:mainfrom
jasro23:kannada_g2p

jasro23 commented Apr 3, 2026

Uh oh!

annagirimokshith commented Apr 5, 2026

Uh oh!

jasro23 commented Apr 9, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 20, 2026

Uh oh!

Copilot AI Apr 20, 2026

Uh oh!

Copilot AI Apr 20, 2026

Uh oh!

Copilot AI Apr 20, 2026

Uh oh!

Copilot AI Apr 20, 2026

Uh oh!

Copilot AI Apr 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

-            phoneme_dict = (
-                self._parse_phoneme_dict(phoneme_dict, phoneme_prefix)
-                if isinstance(phoneme_dict, (str, pathlib.Path))
-                else phoneme_dict
-            )
+            if isinstance(phoneme_dict, (str, pathlib.Path)):
+                phoneme_dict = self._parse_phoneme_dict(phoneme_dict, phoneme_prefix)
+            else:
+                normalized_phoneme_dict = {}
+                for word, prons in phoneme_dict.items():
+                    normalized_prons = []
+                    for pron in prons:
+                        if isinstance(pron, str):
+                            normalized_prons.extend([phoneme_prefix + symbol for symbol in pron])
+                        else:
+                            normalized_prons.extend(pron)
+                    normalized_phoneme_dict[word] = normalized_prons
+                phoneme_dict = normalized_phoneme_dict

Conversation

jasro23 commented Apr 3, 2026

What does this PR do ?

Changelog

Usage

GitHub Actions CI

Before your PR is "Ready for review"

Who can review?

Additional Information

Uh oh!

annagirimokshith commented Apr 5, 2026

Uh oh!

jasro23 commented Apr 9, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants