Expose normalized encoder text

README.md

@@ -57,6 +57,9 @@ tts.save_wav("teratts.wav", waveform)
 `waveform` is a mono `float32` NumPy array at 44,100 Hz. `save_wav` writes
 standard signed-16-bit PCM WAV without an extra audio package.
 
+To inspect the exact text passed to the encoder after number expansion, stress
+marking, and Unicode normalization, call `tts.normalize_text(text)`.
+
 ## Controls
 
 | Control | Values | Effect |

modeling_teratts.py

@@ -18,6 +18,7 @@ from .teratts import (
     generate_speech,
     generate_speech_stream,
     load_model,
+    normalize_text,
     write_wav,
 )
 
@@ -122,6 +123,10 @@ class TeraTTSModel(PreTrainedModel):
             seed=seed,
         )
 
+    def normalize_text(self, text: str) -> str:
+        """Show the final tagged text passed to the TTS text encoder."""
+        return normalize_text(self._runtime(), text)
+
     def generate_speech_stream(
         self,
         text: str,

teratts.py

@@ -101,6 +101,13 @@ def expand_tagged_numbers(text: str) -> str:
     return LANGUAGE_TAG.sub(expand_span, text)
 
 
+def normalize_text(loaded: "LoadedTTS", text: str) -> str:
+    """Return the exact text tensorized by the text encoder for an utterance."""
+    expanded_text = expand_tagged_numbers(text)
+    model_text, _ = prepare_raw_text(add_russian_stress(expanded_text, loaded.accentizer))
+    return model_text
+
+
 class UnicodeIndexer:
     def __init__(self, indexer_path: Path):
         self.table = json.loads(indexer_path.read_text())
@@ -360,10 +367,8 @@ def _generate_latent(
     if style_ttl.shape != (1, 50, 256) or style_dp.shape != (1, 8, 16):
         raise ValueError("style assets have unexpected shapes")
 
-    expanded_text = expand_tagged_numbers(text)
-    model_text, duration_text = prepare_raw_text(
-        add_russian_stress(expanded_text, loaded.accentizer)
-    )
+    model_text = normalize_text(loaded, text)
+    duration_text = model_text.replace("+", "")
     text_ids, text_mask = loaded.indexer.batch(model_text)
     duration_ids, duration_mask = loaded.indexer.batch(duration_text)
     text_emb = loaded.text_encoder.run(