Skip to main content

Text normalization

Use normalize_text() to inspect spoken text before generation.

from kittentts import normalize_text

spoken = normalize_text("Dr. Rivera paid $12.50 at 3:05 p.m.")
print(spoken)

The normalization pipeline handles common English forms including:

  • integers, decimals, scientific notation, and scale suffixes;
  • ordinals, Roman numerals, ranges, fractions, and decades;
  • currency, percentages, units, phone numbers, and IP addresses;
  • times, versions, email addresses, and URLs;
  • abbreviations, punctuation, Unicode, and whitespace.

Track changed spans​

Set return_spans=True to receive normalized text plus original-to-normalized ranges for changed segments.

result = normalize_text("Fig. 2 costs $5", return_spans=True)

print(result.text)
for span in result.spans:
    print(span)

This is useful for editor previews or highlighting what the TTS engine will read differently.

Generate normalized input​

audio = tts.generate(
    "Version 2.1 costs $19.99.",
    voice="Bella",
    clean_text=True,
)

Normalize once. Do not call normalize_text() yourself and then pass clean_text=True unless you intentionally want the pipeline to process the already-normalized string again.

The current locale argument defaults to en-US; the v0.8 models and pipeline are English-focused.

Last updated on