Skip to content

Open-source State-of-the-art TTS model which runs on a CPU 😻

Python

Simplify normalization API

Dewan Shakil7eeae3003f3d
diff --git a/README.md b/README.md
index 0cdc270..6d8bb73 100644
--- a/README.md
+++ b/README.md
@@ -132,8 +132,6 @@ Synthesize speech from text, returning a NumPy array of audio samples at 24 kHz.
| `voice` | `str` | `"expr-voice-5-m"` | Voice name (see available voices) |
| `speed` | `float` | `1.0` | Speech speed multiplier |
| `clean_text` | `bool` | `False` | Preprocess text (expand numbers, currencies, etc.) |
-| `normalize` | `bool` | `None` | Use read-aloud text normalization. When unset, follows `clean_text` for backward compatibility |
-| `locale` | `str` | `"en-US"` | Normalization locale. Currently supports English |
### `model.generate_to_file(text, output_path, voice, speed, sample_rate, clean_text)`
@@ -148,9 +146,9 @@ Synthesize speech and write directly to an audio file.
| `sample_rate` | `int` | `24000` | Audio sample rate in Hz |
| `clean_text` | `bool` | `True` | Preprocess text (expand numbers, currencies, etc.) |
-### `normalize_text(text, locale="en-US", domain="general-read-aloud", return_spans=False)`
+### `normalize_text(text, locale="en-US", return_spans=False)`
-Normalize text for read-aloud use without generating audio.
+Normalize text for TTS without generating audio.
```python
from kittentts import normalize_text
diff --git a/kittentts/get_model.py b/kittentts/get_model.py
index 73c3a37..d8d2225 100644
--- a/kittentts/get_model.py
+++ b/kittentts/get_model.py
@@ -24,12 +24,11 @@ class KittenTTS:
self.model = download_from_huggingface(repo_id=repo_id, cache_dir=cache_dir, backend=backend)
- def normalize_text(self, text, locale="en-US", domain="general-read-aloud", return_spans=False):
- """Normalize text for read-aloud synthesis without generating audio."""
- return normalize_text(text, locale=locale, domain=domain, return_spans=return_spans)
+ def normalize_text(self, text, locale="en-US", return_spans=False):
+ """Normalize text for TTS without generating audio."""
+ return normalize_text(text, locale=locale, return_spans=return_spans)
- def generate(self, text, voice="expr-voice-5-m", speed=1.0, clean_text=False,
- normalize=None, locale="en-US", domain="general-read-aloud"):
+ def generate(self, text, voice="expr-voice-5-m", speed=1.0, clean_text=False):
"""Generate audio from text.
Args:
@@ -41,35 +40,17 @@ class KittenTTS:
Audio data as numpy array
"""
print(f"Generating audio for text: {text}")
- return self.model.generate(
- text,
- voice=voice,
- speed=speed,
- clean_text=clean_text,
- normalize=normalize,
- locale=locale,
- domain=domain,
- )
+ return self.model.generate(text, voice=voice, speed=speed, clean_text=clean_text)
- def generate_stream(self, text, voice="expr-voice-5-m", speed=1.0, clean_text=False,
- normalize=None, locale="en-US", domain="general-read-aloud"):
+ def generate_stream(self, text, voice="expr-voice-5-m", speed=1.0, clean_text=False):
"""Generate audio as a stream of chunks.
Yields:
numpy.ndarray: Audio data for each text chunk.
"""
- yield from self.model.generate_stream(
- text,
- voice=voice,
- speed=speed,
- clean_text=clean_text,
- normalize=normalize,
- locale=locale,
- domain=domain,
- )
+ yield from self.model.generate_stream(text, voice=voice, speed=speed, clean_text=clean_text)
- def generate_to_file(self, text, output_path, voice="expr-voice-5-m", speed=1.0, sample_rate=24000,
- clean_text=True, normalize=None, locale="en-US", domain="general-read-aloud"):
+ def generate_to_file(self, text, output_path, voice="expr-voice-5-m", speed=1.0, sample_rate=24000):
"""Generate audio from text and save to file.
Args:
@@ -79,17 +60,7 @@ class KittenTTS:
speed: Speech speed (1.0 = normal)
sample_rate: Audio sample rate
"""
- return self.model.generate_to_file(
- text,
- output_path,
- voice=voice,
- speed=speed,
- sample_rate=sample_rate,
- clean_text=clean_text,
- normalize=normalize,
- locale=locale,
- domain=domain,
- )
+ return self.model.generate_to_file(text, output_path, voice=voice, speed=speed, sample_rate=sample_rate)
@property
def available_voices(self):
diff --git a/kittentts/onnx_model.py b/kittentts/onnx_model.py
index 09138b8..b368e4f 100644
--- a/kittentts/onnx_model.py
+++ b/kittentts/onnx_model.py
@@ -114,29 +114,25 @@ class KittenTTS_1_Onnx:
"speed": np.array([speed], dtype=np.float32),
}
- def normalize_text(self, text: str, locale: str = "en-US", domain: str = "general-read-aloud", return_spans: bool = False):
- return normalize_text(text, locale=locale, domain=domain, return_spans=return_spans)
+ def normalize_text(self, text: str, locale: str = "en-US", return_spans: bool = False):
+ return normalize_text(text, locale=locale, return_spans=return_spans)
- def generate(self, text: str, voice: str = "expr-voice-5-m", speed: float = 1.0, clean_text: bool=True,
- normalize: bool = None, locale: str = "en-US", domain: str = "general-read-aloud") -> np.ndarray:
+ def generate(self, text: str, voice: str = "expr-voice-5-m", speed: float = 1.0, clean_text: bool=True) -> np.ndarray:
out_chunks = []
- should_normalize = clean_text if normalize is None else normalize
- if should_normalize:
- text = normalize_text(text, locale=locale, domain=domain)
+ if clean_text:
+ text = self.preprocessor(text)
for text_chunk in chunk_text(text):
out_chunks.append(self.generate_single_chunk(text_chunk, voice, speed))
return np.concatenate(out_chunks, axis=-1)
- def generate_stream(self, text: str, voice: str = "expr-voice-5-m", speed: float = 1.0, clean_text: bool = True,
- normalize: bool = None, locale: str = "en-US", domain: str = "general-read-aloud"):
+ def generate_stream(self, text: str, voice: str = "expr-voice-5-m", speed: float = 1.0, clean_text: bool = True):
"""Generate audio chunk-by-chunk as a generator.
Yields:
numpy.ndarray: Audio data for each text chunk.
"""
- should_normalize = clean_text if normalize is None else normalize
- if should_normalize:
- text = normalize_text(text, locale=locale, domain=domain)
+ if clean_text:
+ text = self.preprocessor(text)
for text_chunk in chunk_text(text):
yield self.generate_single_chunk(text_chunk, voice, speed)
@@ -161,8 +157,7 @@ class KittenTTS_1_Onnx:
return audio
def generate_to_file(self, text: str, output_path: str, voice: str = "expr-voice-5-m",
- speed: float = 1.0, sample_rate: int = 24000, clean_text: bool=True,
- normalize: bool = None, locale: str = "en-US", domain: str = "general-read-aloud") -> None:
+ speed: float = 1.0, sample_rate: int = 24000, clean_text: bool=True) -> None:
"""Synthesize speech and save to file.
Args:
@@ -173,6 +168,6 @@ class KittenTTS_1_Onnx:
sample_rate: Audio sample rate
clean_text: If true, it will cleanup the text. Eg. replace numbers with words.
"""
- audio = self.generate(text, voice, speed, clean_text=clean_text, normalize=normalize, locale=locale, domain=domain)
+ audio = self.generate(text, voice, speed, clean_text=clean_text)
sf.write(output_path, audio, sample_rate)
print(f"Audio saved to {output_path}")
diff --git a/kittentts/preprocess.py b/kittentts/preprocess.py
index 56700dc..b31ed00 100644
--- a/kittentts/preprocess.py
+++ b/kittentts/preprocess.py
@@ -650,7 +650,7 @@ class NormalizedSpan:
@dataclass
class NormalizedTextResult:
- """Normalized read-aloud text plus optional original-to-normalized spans."""
+ """Normalized text plus optional original-to-normalized spans."""
text: str
spans: List[NormalizedSpan]
@@ -916,31 +916,26 @@ def _sub_with_spans(
def normalize_text(
text: str,
locale: str = "en-US",
- domain: str = "general-read-aloud",
return_spans: bool = False,
) -> Union[str, NormalizedTextResult]:
- """Normalize English text for read-aloud TTS use cases.
+ """Normalize English text for TTS use cases.
Args:
text: Input text to normalize.
locale: Currently only "en-US" is supported.
- domain: Currently only "general-read-aloud" is supported.
return_spans: When true, return NormalizedTextResult instead of text.
"""
- result = normalize_text_result(text, locale=locale, domain=domain)
+ result = normalize_text_result(text, locale=locale)
return result if return_spans else result.text
def normalize_text_result(
text: str,
locale: str = "en-US",
- domain: str = "general-read-aloud",
) -> NormalizedTextResult:
"""Normalize English text and return span metadata for changed segments."""
if locale.lower() not in {"en-us", "en"}:
raise ValueError("Only en-US text normalization is currently supported")
- if domain != "general-read-aloud":
- raise ValueError("Only the general-read-aloud normalization domain is currently supported")
text = normalize_unicode(text)
origins = list(range(len(text)))
diff --git a/tests/test_text_normalization.py b/tests/test_text_normalization.py
index 9b1487b..6a9eb92 100644
--- a/tests/test_text_normalization.py
+++ b/tests/test_text_normalization.py
@@ -5,7 +5,7 @@ from kittentts.preprocess import chunk_text
class TextNormalizationTests(unittest.TestCase):
- def test_issue_examples_normalize_for_read_aloud(self):
+ def test_issue_examples_normalize_for_tts(self):
cases = {
"Smith et al. 2024, pp. 31-35": "Smith et al twenty twenty-four pages thirty-one to thirty-five",
"Fig. 2": "Figure two",