integrated expand date function into the preprocessing function

[AdiistheGoat]

Summary

This PR adds numeric date expansion to the preprocessing pipeline so common date formats are converted into a natural spoken form before inference. This prevents the model from reading dates character-by-character (e.g., “zero three slash…”) and improves pronunciation.

Audio comparison

• test_date_raw.wav: output with the current preprocessing script
• test_date.wav: output with the updated preprocessing script

Attachments:

• test_date_raw.wav
• test_date.wav

---

Changes

1) New regex patterns (around lines ~215–220)

Added three standalone date matchers (with lookbehind/lookahead to avoid matching parts of larger numbers):

• _RE_DATE_slash: slash-separated dates (D/M/YY, DD/MM/YYYY)
• _RE_DATE_hyphen: hyphen-separated dates (D-M-YY, DD-MM-YYYY)
• _RE_DATE_ISO: ISO-style dates (YYYY-M-D, YYYY-MM-DD)

2) New expand_dates(text) helper

Introduced expand_dates(text) to normalize numeric date formats into a spoken, human-readable form.

Examples

• 03/14/2022 → 14th March twenty twenty-two
• 2022-03-14 → 14th March twenty twenty-two

Supported formats

• Slash style: D/M/YY, DD/MM/YYYY (also handles M/D/... where applicable)
• Hyphen style: D-M-YY, DD-MM-YYYY
• ISO style: YYYY-MM-DD

Day/month resolution (ambiguity handling)
Numeric dates like 03/04/2022 are ambiguous. Resolution rules:

• If one of the first two components is > 12, that component must be the day
• If both are <= 12, default to DD/MM

3) Integrated into the pipeline (around line ~910)

Added expand_dates(text) at the beginning of TextPreprocessor.process() so dates are expanded before other transformations.

---

Validation & safety behavior

To avoid accidental or unsafe rewrites, the function applies basic validation before converting:

• If day == 0 or month == 0, leave the original substring unchanged
• Range checks:
  • day <= 31
  • month <= 12
• No full calendar validation (e.g., 30/02/2022 is not rejected at the calendar level)
• If a match fails validation, the original text is preserved unchanged

---

Year handling

• Two-digit years are normalized as 2000 + year
  • Example: 22 → 2022
• The normalized year is converted to words via the existing number_to_words() helper

---

Output format

All supported inputs normalize to the canonical spoken form:

{ordinal_day} {MonthName} {year_in_words}

Example:

• 14/03/2022 → 14th March twenty twenty-two

Supporting mappings introduced:

• day_mappings: day number → ordinal (e.g., 1 -> 1st, 2 -> 2nd, …)
• month_mappings: month number → month name (e.g., 1 -> January, 2 -> February, …)

---

Why this matters

This update makes text more TTS-friendly by converting numeric date formats into natural language, improving intelligibility and overall output quality.

---

Date Expansion Examples

The expand_dates(text) function converts supported numeric date formats into a spoken form.

Common cases

Input	Output
03/14/2022	14th March two thousand twenty-two
2022-03-14	14th March two thousand twenty-two
14/03/2022	14th March two thousand twenty-two
14-03-2022	14th March two thousand twenty-two
03/14/22	14th March two thousand twenty-two
2022-03-21	21st March two thousand twenty-two

Ambiguous date handling

When both the first and second numeric parts are <= 12, the function defaults to DD/MM.

Input	Output
03/04/2022	3rd April two thousand twenty-two
05/04/2022	5th April two thousand twenty-two
04/05/2022	4th May two thousand twenty-two
01/02/03	1st February two thousand three

Mixed format support

Input	Output
7/8/22	7th August two thousand twenty-two
08/07/2022	8th July two thousand twenty-two
2022-08-07	7th August two thousand twenty-two
03-17-2022	17th March two thousand twenty-two

Invalid or protected cases

These cases are intentionally left unchanged when they should not be treated as dates or do not pass validation.

Input	Output
192.168.03.14	192.168.03.14
03/14	03/14
00/12/2022	00/12/2022
12/00/2022	12/00/2022
2022-00-10	2022-00-10
03 / 14 / 2022	03 / 14 / 2022
2022 - 03 - 14	2022 - 03 - 14

Edge cases

Input	Output
02/30/2022	30th February two thousand twenty-two
02/29/2024	29th February two thousand twenty-four
02/29/2023	29th February two thousand twenty-three
03/14/99	14th March two thousand ninety-nine
03/14/00	14th March two thousand
03/14/68	14th March two thousand sixty-eight
run-2022-03-14-0007	run-14th March two thousand twenty-two-0007

[Krystal5222]

What does all this mean I am so clueless

…

On Mar 9, 2026 1:01 AM, "Aditya Goyal" ***@***.***> wrote: test_date_raw.wav <https://github.com/user-attachments/files/25832597/test_date_raw.wav> test_date.wav <https://github.com/user-attachments/files/25832598/test_date.wav> test_date_raw.wav is the output with the current preprocessing script test_date.wav is the output with the updated preprocessing script Added Date Expansion Functionality 1. New Regex Patterns (around line 215–220) Added three regex patterns to match different date formats: • _RE_DATE_slash: Matches slash-separated dates (D/M/YY or DD/MM/YYYY) • _RE_DATE_hyphen: Matches hyphen-separated dates (D-M-YY or DD-MM-YYYY) • _RE_DATE_ISO: Matches ISO format dates (YYYY-M-D or YYYY-MM-DD) All patterns use lookbehind/lookahead assertions to ensure dates are standalone (not part of larger numbers). 2. New expand_dates() Function (around line 700–800) Added a date expansion function that: • Converts numeric dates to spoken words (e.g., "03/14/2022" → "14th March twenty twenty-two") • Handles multiple formats: slash dates, hyphen dates, and ISO dates • Validates dates: day (1–31), month (1–12), and year ranges • Resolves ambiguity: when both parts are ≤12, defaults to DD/MM; when one part is >12, treats it as the day • Normalizes 2-digit years: adds 2000 to years <100 (e.g., "22" → "2022") • Preserves invalid dates by returning the original substring if validation fails • Uses internal day_mappings (ordinal suffixes) and month_mappings (month names) • Leverages existing number_to_words() for converting the year to words 3. Integrated into Pipeline (around line 910) Added an expand_dates(text) call at the beginning of TextPreprocessor.process(), ensuring dates are expanded before other transformations. Summary This update introduces date normalization that converts numeric date formats into natural, spoken language—making the text more suitable for TTS output. ------------------------------ You can view, comment on, or merge this pull request online at: #107 Commit Summary - cfc5653 <cfc5653> integrated expand date function File Changes (1 file <https://github.com/KittenML/KittenTTS/pull/107/files>) - *M* kittentts/preprocess.py <https://github.com/KittenML/KittenTTS/pull/107/files#diff-2cde1eea64f167aa37f9d061f653f4ea01b0f6b644e2c7193141953e887b09d7> (116) Patch Links: - https://github.com/KittenML/KittenTTS/pull/107.patch - https://github.com/KittenML/KittenTTS/pull/107.diff — Reply to this email directly, view it on GitHub <#107>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A5AYBKCOJJOWQVA7YFRJIJ34PZM5DAVCNFSM6AAAAACWLNC5BWVHI2DSMVQWIX3LMV43ASLTON2WKOZUGA2DGMZWGA3DENI> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[AdiistheGoat]

Hey @Krystal5222, thanks for raising this. I may not have explained it clearly enough in the PR description, but I’m happy to update it later today to make it more readable and add more examples as well. In the meantime, the class comments may also provide some additional context. Let me know if there’s anything specific you’d like me to clarify.

Edit: I have improved the PR description. Lmk if something doesn't make sense.