danielrosehill's picture
Seed v0.5: Israeli names and frames for Israeli addresses
85bed5a verified
|
Raw History Blame Contribute Delete
5.83 kB
metadata
license: cc-by-4.0
language:
  - en
task_categories:
  - text-generation
  - token-classification
pretty_name: ASR Formatter Training Pairs
size_categories:
  - 10K<n<100K
tags:
  - inverse-text-normalization
  - asr-post-processing
  - email
  - synthetic
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.jsonl
      - split: validation
        path: validation.jsonl
      - split: test
        path: test.jsonl
      - split: recorder
        path: recorder.jsonl

ASR Formatter Training Pairs

Text pairs for a small formatter model that sits between a speech recognition model and any optional clean-up model. The formatter turns spoken forms that an ASR tokenizer cannot write into written text, and it leaves everything else alone.

Why it exists: the Parakeet TDT 0.6B v2 tokenizer has no @, _ or +, so no Parakeet fine-tune can write an email address. The Hebrish fine-tunes are trained to write a fixed spoken form instead, and the formatter writes the address:

input (what the ASR writes) target (what the formatter writes)
Email dana underscore levi at tidewell.io with any questions. Email dana_levi@tidewell.io with any questions.
Write noa dot cohen at gmail dot com on the form. Write noa.cohen@gmail.com on the form.
I'll meet you at the station at five. I'll meet you at the station at five. (unchanged)

Seed v0.5 (2026-11-09): email addresses only. v0.5 makes Israeli addresses Israeli throughout: every address on an Israeli domain and every Israeli kind uses Israeli names (diaspora Jewish and immigrant names for oleh-webmail, with their own held-out test pools), and half of them sit in Israeli-life sentences (arnona, va'ad bayit, gan, Bituach Leumi, tofes 106). v0.4 added the public and ISP mail of the countries immigrants to Israel come from, the older addresses middle-aged and older people still use (kind oleh-webmail: aol.com, comcast.net, verizon.net, att.net, btinternet.com, sky.com, rogers.com, sympatico.ca, bigpond.com, mweb.co.za, orange.fr, mail.ru, yandex.ru, uol.com.br and others, weighted US first), country TLDs for those countries, and cuts Irish domains (.ie, eircom.net, live.ie) to one entry each (about 1% of addresses). v0.3 added an Israeli quarter of the positives: businesses on Gmail (business-gmail, the commonest), old ISP and portal mail (israeli-isp: bezeqint.net, bezeq.net, netvision.net.il, 012.net.il, zahav.net.il, walla, 013 and others), Gmail locals with phone digits (il-name-phone), university mail (academic-il), more municipalities and ministries (municipal, gov-il, plus national bodies on .org.il and idf.il); Proton (protonmail.com, pm.me) and more weight on gmail.com; and CC / BCC instructions, with bare "CC me" / "BCC me" rows as negatives. Previously: v0.1 added "hyphen" beside "dash" in the spoken-out style; v0.2 adds newer TLDs (.ai, .cloud, .app and others), Israeli municipal and government domains (<city>.muni.il, <city>.gov.il, <ministry>.gov.il; kinds municipal, gov-il) and the maintainer's two public addresses (kind own-address, train only). More scope (underscored identifiers like Q4_K_M, degree signs, plus signs) will be added as new scope values.

Columns

Column Meaning
id fmt_ + hash of source, input and target
input text as an ASR model writes it; only characters the Parakeet v2 tokenizer has
target the written form
changed input != target
scope email, or none for rows that must pass through unchanged
kind the address pattern: role, role-compound, role-underscore, first, firstlast, first.last, first_last, first-last, initial-last, initial.last, initial_last, first.initial, last.first, last_first, name-digits, handle, plus-tag, hyphen-domain, digit-domain, subdomain, country-domain, webmail, municipal, gov-il, business-gmail, israeli-isp, il-name-phone, academic-il, oleh-webmail, own-address; none for negatives
input_style canonical (the fine-tunes' spoken form: local part joined, . and - written, _ "underscore", + "plus", " at ", domain as typed), dots-spoken (every dot said, hyphens as "dash" or "hyphen", as stock models write it), mixed (domain dots spoken), capitalised (names capitalised; the target is lowercase), none
n_addresses 0, 1 or 2
source synthetic-email, synthetic-splice (an address clause added to a real dictation sentence), bank-negative (dictation sentences, unchanged), synthetic-negative (ordinary uses of at / plus / dot / underscore and bare domains), recorder-prompt
split as below

Splits

Split Rows What
train 33,340 synthetic emails, splices and negatives
validation 1,722 1 in 20 of the same pools, by hash
test 1,500 addresses built from held-out first names, surnames and company names never seen in train
recorder 542 the prompts of the Hebrish capture app task email-addresses: the reference for scoring real recordings, decoded by an ASR model and then formatted (audio: Hebrish/Email-Address-Sentences)

25,240 rows change and 12,030 are negatives.

Privacy

Every address is invented except the maintainer's own two public addresses (public@ and daniel@danielrosehill.com): names come from a mixed Israeli, Irish and international pool, company domains are made up, and webmail domains (gmail.com, outlook.com and others) carry invented local parts. Negatives are dictation-practice sentences from the Hebrish sentence bank, with personal and medical categories excluded.

Build

formatter/build_pairs.py in the Hebrish training repo (seeded, deterministic, stdlib only).