[Bugfix] Allow Fish Speech phoneme-control tokens through normalize_fish_speech_text (#7043)

Signed-off-by: Pujitha Paladugu <10557236+pujitha24@users.noreply.github.com>
Co-authored-by: Pujitha Paladugu <10557236+pujitha24@users.noreply.github.com>

pyproject.toml

@@ -376,6 +376,9 @@ ba = "ba"
 SER = "SER"
 # NumPy ndarray.flags.writeable (canonical flag name, not "writable").
 writeable = "writeable"
+# ARPAbet phoneme symbol (Fish Speech phoneme-control token example, e.g.
+# "IY0"), not a typo of "IT".
+IY = "IY"
 # "convnet" abbreviates convolutional network. It appears in the vendored Fish
 # Speech DAC codec identifiers (convnet_type, transconvnet_type) and is not a
 # typo of "convent".

tests/model_executor/models/fish_speech/test_prompt_utils.py

@@ -0,0 +1,23 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+import pytest
+
+from vllm_omni.model_executor.models.fish_speech.prompt_utils import normalize_fish_speech_text
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+
+def test_normalize_fish_speech_text_allows_phoneme_control_tokens():
+    text = "Deploy with <|phoneme_start|>K UW2 B ER0 N EH1 T IY0 Z<|phoneme_end|>."
+
+    assert normalize_fish_speech_text(text) == text
+
+
+def test_normalize_fish_speech_text_rejects_unknown_control_token():
+    with pytest.raises(ValueError, match=r"unsupported control token\(s\): <\|foo\|>"):
+        normalize_fish_speech_text("<|foo|>hello")
+
+
+def test_normalize_fish_speech_text_normalizes_legacy_speaker_tag():
+    assert normalize_fish_speech_text("<speaker:0>hello") == "<|speaker:0|>hello"

vllm_omni/model_executor/models/fish_speech/prompt_utils.py

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """Shared Fish Speech prompt construction helpers."""
 
 from __future__ import annotations
@@ -11,6 +14,7 @@ FISH_CLONE_SYSTEM_PROMPT_SUFFIX = "\n\nSpeech:\n"
 
 _LEGACY_SPEAKER_TAG_PATTERN = re.compile(r"<speaker:(\d+)>")
 _CANONICAL_SPEAKER_TAG_PATTERN = re.compile(r"<\|speaker:\d+\|>")
+_PHONEME_TAG_PATTERN = re.compile(r"<\|phoneme_(?:start|end)\|>")
 _CONTROL_TOKEN_PATTERN = re.compile(r"<\|[^>]+\|>")
 
 
@@ -21,7 +25,7 @@ def normalize_fish_speech_text(text: str, *, add_default_speaker: bool = False)
     disallowed_tokens = [
         token
         for token in _CONTROL_TOKEN_PATTERN.findall(normalized)
-        if not _CANONICAL_SPEAKER_TAG_PATTERN.fullmatch(token)
+        if not (_CANONICAL_SPEAKER_TAG_PATTERN.fullmatch(token) or _PHONEME_TAG_PATTERN.fullmatch(token))
     ]
     if disallowed_tokens:
         disallowed_list = ", ".join(sorted(set(disallowed_tokens)))