Skip to content

feat: English/Spanish call ASR normalizations for gladia-3 - #32

Open
egenthon-cmd wants to merge 1 commit into
mainfrom
feat/en-es-asr-normalizations
Open

egenthon-cmd wants to merge 1 commit into
mainfrom
feat/en-es-asr-normalizations

Conversation

@egenthon-cmd

Copy link
Copy Markdown
Contributor

What does this PR do?

Adds English and Spanish ASR/WER canonicalizations from jail-call mismatches (word + sentence replacements), and tightens English digit+multiplier / digit+point handling so forms like 25 hundred and 44 for point 5 align under gladia-3.

Type of change

  • Edit existing language (fix a replacement, tweak config, …)
  • Edit existing step (bug fix, behaviour change)

Checklist

Edit existing language

  • New/changed word substitutions go in replacements.py, not inline in operators.py
  • If you changed a config field that can be None: the step reading it still handles None gracefully
  • Unit tests updated or added
  • E2e CSV updated if the expected output changed

Edit existing step

  • Step name is unchanged — if the output changes, create a new step name + new preset instead
  • No language-specific logic or string literals added inside the step
  • Unit tests updated or added
  • If the docstring changed: ran uv run scripts/generate_step_docs.py

How was this tested?

uv run pytest tests/e2e/normalization_test.py -k 'gladia_3_en or gladia_3_es'

208 passed.

Summary

  • English: jail-call word/sentence replacements (prepai→prepaid, because→cuz, eleventh 2000 and twelve→112012, …)
  • Spanish: ahorita→ahora, llama→llame, cuchillo→culchi y yo
  • Number normalizer: N hundred / N thousand and N point M digit forms

Made with Cursor

Canonicalize jail-call WER mismatches (homophones, slang, number forms)
so hyp and ref align under gladia-3.

Co-authored-by: Cursor <cursoragent@cursor.com>
@jqueguiner jqueguiner changed the title feat: English/Spanish jail-call ASR normalizations for gladia-3 feat: English/Spanish call ASR normalizations for gladia-3 Sep 28, 2026
paul,powell
streets,street
because,cuz
prepai,prepaid

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

two billion people,2000000000 people
itchys,itches
placed,placer
ruby,rudy

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

three thousand five hundred,3500
two billion people,2000000000 people
itchys,itches
placed,placer

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

duckets,duckers
raytwine,raytron
twine,twan
paul,powell

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

pound,pounds
crystals,pistols
duckets,duckers
raytwine,raytron

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

fimo,famo
holla,holler
pound,pounds
crystals,pistols

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

shi,shit
rudolph,rudolf
fimo,famo
holla,holler

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

talk,talked
shi,shit
rudolph,rudolf
fimo,famo

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

peoples,people
dated,date
telephone,cell phone
try,tried

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

ruby,rudy
cannot,could not
peoples,people
dated,date

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

telephone,cell phone
try,tried
talk,talked
shi,shit

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

holla,holler
pound,pounds
crystals,pistols
duckets,duckers

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

crystals,pistols
duckets,duckers
raytwine,raytron
twine,twan

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

raytwine,raytron
twine,twan
paul,powell
streets,street

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

all right,alright
fire fire,firefires
paul s,pauls
ray tuan,raytron

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

fire fire,firefires
paul s,pauls
ray tuan,raytron
pre prepa,prepaid

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not this one

ray tuan,raytron
pre prepa,prepaid
30 two,32
44 for point 5,44 point 5

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

shoudl be done with a words2num2 normalization

eleventh 2000 and twelve,112012
eleventh 2 thousand and twelve,112012
eleventh 2 thousaond and twelve,112012
25 hundred,2500

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

shoudl be done with a words2num2 normalization

44 for point 5,44 point 5
eleventh 2000 and twelve,112012
eleventh 2 thousand and twelve,112012
eleventh 2 thousaond and twelve,112012

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

shoudl be done with a words2num2 normalization

30 two,32
44 for point 5,44 point 5
eleventh 2000 and twelve,112012
eleventh 2 thousand and twelve,112012

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

shoudl be done with a words2num2 normalization

pre prepa,prepaid
30 two,32
44 for point 5,44 point 5
eleventh 2000 and twelve,112012

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

shoudl be done with a words2num2 normalization

eleventh 2 thousand and twelve,112012
eleventh 2 thousaond and twelve,112012
25 hundred,2500
35 hundred,3500

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

shoudl be done with a words2num2 normalization

eleventh 2 thousaond and twelve,112012
25 hundred,2500
35 hundred,3500
20 thirteen,2013

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

shoudl be done with a words2num2 normalization

quinientos,500
quince mil,15000
ahorita,ahora
llama,llame

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not the same

quince mil,15000
ahorita,ahora
llama,llame
cuchillo,culchi y yo

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

no

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants