feat: rename model

.gitattributes

@@ -11,6 +11,7 @@
 *.mlmodel filter=lfs diff=lfs merge=lfs -text
 *.model filter=lfs diff=lfs merge=lfs -text
 *.msgpack filter=lfs diff=lfs merge=lfs -text
+*.nemo filter=lfs diff=lfs merge=lfs -text
 *.npy filter=lfs diff=lfs merge=lfs -text
 *.npz filter=lfs diff=lfs merge=lfs -text
 *.onnx filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1 @@
+ASR/whisper/training_metrics.png

ASR/parakeet/README.md

@@ -0,0 +1,115 @@
+---
+language:
+  - en
+license: other
+tags:
+  - nemo
+  - parakeet
+  - tdt
+  - automatic-speech-recognition
+  - air-traffic-control
+  - atc
+  - singapore
+  - military
+base_model: nvidia/parakeet-tdt-0.6b-v2
+pipeline_tag: automatic-speech-recognition
+metrics:
+  - wer
+model-index:
+  - name: parakeet-tdt-0.6b-v2-atc-singapore
+    results:
+      - task:
+          type: automatic-speech-recognition
+        metrics:
+          - name: Validation WER
+            type: wer
+            value: 1.14
+---
+
+# Parakeet-TDT 0.6B v2 - Singapore Military ATC
+
+Fine-tuned NVIDIA Parakeet-TDT 0.6B v2 for Singapore Air Force air traffic control speech recognition.
+
+## Performance
+
+| Checkpoint | Validation WER | Notes |
+|------------|----------------|-------|
+| `model.ckpt` | **1.14%** | Best checkpoint |
+| `epoch=54-val_wer=0.0245-last.ckpt` | 2.45% | Final checkpoint, not published here |
+
+## Model Details
+
+| Key | Value |
+|-----|-------|
+| Base model | `nvidia/parakeet-tdt-0.6b-v2` |
+| Framework | NeMo |
+| Model class | `EncDecRNNTBPEModel` |
+| Format | Raw `.ckpt` checkpoint + tokenizer artifacts |
+| Checkpoint size | 7.0 GB |
+| Domain | Singapore military ATC (Tengah WSAT, Paya Lebar WSAP) |
+
+## Included Files
+
+- `model.ckpt` - best fine-tuned checkpoint
+- `artifacts/705f11d22dc04b169effc35ce5cd1361_tokenizer.model`
+- `artifacts/a4715c7f6b2d4c2bb709306073d0c0a4_tokenizer.vocab`
+- `artifacts/4cf78c8ca4ca44fca36c3754478fb188_vocab.txt`
+
+## Usage
+
+This repo currently publishes the fine-tuned model as a raw NeMo checkpoint rather than a packaged `.nemo` archive. The tokenizer artifact paths therefore need to be pointed at the local `artifacts/` folder before restore.
+
+```python
+from pathlib import Path
+
+import torch
+from omegaconf import OmegaConf
+from nemo.collections.asr.models import EncDecRNNTBPEModel
+
+model_dir = Path("ASR/parakeet")
+ckpt_path = model_dir / "model.ckpt"
+artifacts_dir = model_dir / "artifacts"
+
+bundle = torch.load(ckpt_path, map_location="cpu", weights_only=False)
+cfg = bundle["hyper_parameters"]["cfg"]
+cfg = OmegaConf.create(OmegaConf.to_container(cfg, resolve=False))
+cfg.tokenizer.model_path = str(
+    artifacts_dir / "705f11d22dc04b169effc35ce5cd1361_tokenizer.model"
+)
+cfg.tokenizer.vocab_path = str(
+    artifacts_dir / "4cf78c8ca4ca44fca36c3754478fb188_vocab.txt"
+)
+cfg.tokenizer.spe_tokenizer_vocab = str(
+    artifacts_dir / "a4715c7f6b2d4c2bb709306073d0c0a4_tokenizer.vocab"
+)
+
+model = EncDecRNNTBPEModel.load_from_checkpoint(str(ckpt_path), cfg=cfg)
+model.eval().cuda()
+
+hypotheses = model.transcribe(
+    ["audio.wav"],
+    return_hypotheses=True,
+    timestamps=True,
+)
+
+if isinstance(hypotheses, tuple):
+    hypotheses = hypotheses[0]
+
+hyp = hypotheses[0]
+print(hyp.text)
+print(hyp.timestamp["word"])
+```
+
+Tested in ASTRA with `nemo_toolkit[asr]==2.7.3`.
+
+## Output Format
+
+The model outputs normalized spoken text intended for downstream ATC formatting:
+
+| Input audio says | Model outputs |
+|-----------------|---------------|
+| "CAMEL climb flight level zero nine zero" | `camel climb flight level zero nine zero` |
+| "Contact Tengah Approach one three zero decimal zero" | `contact tengah approach one three zero decimal zero` |
+| "Squawk four five two one" | `squawk four five two one` |
+
+ASTRA then applies a deterministic formatter to convert normalized speech into display text such as `CAMEL climb FL090`.

ASR/parakeet/artifacts/4cf78c8ca4ca44fca36c3754478fb188_vocab.txt

@@ -0,0 +1,1023 @@
+t
+th
+a
+##in
+the
+##re
+w
+o
+s
+##at
+##ou
+##er
+##nd
+i
+b
+c
+##on
+h
+##ing
+to
+m
+##en
+f
+p
+##an
+d
+##es
+##or
+##ll
+of
+and
+y
+l
+I
+##it
+in
+##is
+##ed
+g
+you
+##ar
+that
+##om
+##as
+n
+##ve
+##us
+##ic
+##ow
+##al
+it
+be
+wh
+##le
+##ion
+##ut
+##ot
+we
+is
+e
+##et
+##ay
+re
+on
+T
+A
+ha
+##ent
+##ke
+##ct
+S
+##ig
+##ver
+Th
+##all
+##id
+for
+##ro
+he
+##se
+this
+##ld
+##ly
+go
+k
+st
+##st
+##ch
+li
+u
+##am
+##ur
+##ce
+##ith
+##im
+so
+have
+do
+##ht
+##th
+an
+with
+##ad
+r
+##ir
+was
+as
+W
+are
+##ust
+##ally
+j
+se
+##ation
+##od
+##ere
+like
+not
+kn
+##ight
+B
+they
+And
+know
+##ome
+##op
+can
+or
+sh
+me
+##ill
+##ant
+##ck
+what
+at
+ab
+##ould
+##ol
+So
+C
+##use
+##ter
+##il
+but
+just
+ne
+de
+##ra
+##ore
+there
+##ul
+##out
+con
+all
+The
+##ers
+H
+fr
+pro
+##ge
+##ea
+Y
+O
+M
+##pp
+com
+##ess
+ch
+al
+##est
+##ate
+##qu
+lo
+ex
+##very
+su
+##ain
+one
+##ca
+##art
+##ist
+##if
+##ive
+if
+##ink
+##nt
+##ab
+about
+going
+v
+wor
+##um
+##ok
+your
+my
+##ind
+get
+##cause
+from
+don
+##ri
+##pe
+##un
+##ity
+up
+P
+out
+##ort
+L
+##ment
+##el
+N
+some
+##ich
+##and
+think
+##em
+##oug
+G
+##os
+D
+##res
+because
+by
+##ake
+int
+##ie
+us
+tr
+then
+##ack
+pl
+here
+pe
+##her
+will
+F
+which
+##ard
+right
+thing
+want
+##ies
+##ople
+It
+them
+##ame
+We
+##our
+say
+R
+people
+see
+who
+##ast
+##ure
+##ect
+##ear
+tim
+E
+You
+would
+when
+##ven
+our
+##ci
+really
+more
+##ound
+##ose
+##ak
+co
+##ide
+##ough
+had
+##so
+qu
+##eah
+were
+##ine
+act
+##ther
+these
+how
+now
+sa
+##ud
+Wh
+man
+##ous
+##one
+##pt
+##ff
+##ong
+has
+any
+very
+But
+look
+##iv
+##itt
+time
+mo
+ar
+##hing
+le
+work
+their
+##are
+his
+##per
+##ions
+im
+ag
+J
+no
+en
+got
+##ag
+sp
+##ans
+##act
+te
+also
+##iz
+##ice
+That
+cl
+been
+way
+fe
+did
+##ple
+##ually
+other
+U
+##ite
+##age
+##omet
+##ber
+##reat
+##ree
+into
+##own
+tw
+part
+##alk
+where
+need
+every
+##pl
+ad
+##ry
+over
+##ble
+##ap
+##ue
+kind
+po
+back
+cont
+##iff
+somet
+pr
+##nder
+##ire
+good
+than
+##ace
+gu
+##ep
+##og
+##ick
+##way
+lot
+un
+things
+In
+##ish
+##kay
+well
+could
+pre
+two
+##irst
+diff
+##ach
+##cc
+##ittle
+##int
+He
+those
+##ence
+##ip
+##ase
+him
+make
+little
+##ical
+gr
+year
+##ass
+thr
+##uch
+##ated
+This
+off
+res
+##ac
+##ance
+actually
+talk
+##ult
+##able
+##orm
+dis
+first
+##ations
+something
+she
+##sel
+let
+##ord
+may
+##ia
+am
+her
+said
+bo
+##be
+##ount
+much
+per
+even
+differe
+##vel
+##ary
+app
+##ving
+comm
+imp
+##ys
+again
+##ress
+yeah
+down
+##ang
+mean
+##na
+##ens
+does
+fo
+comp
+ro
+bl
+##ody
+K
+through
+start
+##uct
+only
+bet
+under
+br
+take
+##ning
+bu
+use
+Ch
+##xt
+##co
+##ory
+##ild
+put
+call
+new
+##other
+##ting
+happ
+##ater
+inc
+##ition
+different
+should
+##ade
+##ign
+##thing
+day
+##fore
+Yeah
+##ark
+##ile
+##ial
+come
+They
+being
+try
+##ious
+sc
+bit
+spe
+##ub
+##fe
+doing
+St
+##vers
+##av
+##ty
+##ian
+##onna
+##red
+##wn
+ke
+##form
+##ors
+fl
+##fter
+##ail
+##ents
+gonna
+point
+##ces
+There
+##self
+many
+If
+same
+sy
+quest
+most
+great
+What
+fu
+##ug
+show
+##we
+##ual
+##ons
+Be
+##ically
+ser
+rem
+ind
+pers
+V
+##he
+str
+##ved
+still
+##ank
+rec
+wr
+##ought
+##day
+##ath
+end
+bas
+##ft
+##erm
+##body
+##ph
+##ject
+##ict
+play
+Is
+##ates
+ph
+##oth
+acc
+##get
+years
+em
+id
+Oh
+##ves
+##ever
+inter
+rel
+before
+feel
+##igh
+three
+##iss
+des
+##ne
+why
+uh
+To
+cons
+hel
+after
+##ower
+##urn
+okay
+long
+bel
+around
+##ful
+##te
+##ise
+ob
+supp
+##ady
+##ange
+##aking
+pos
+##atch
+tra
+##gr
+might
+##ert
+help
+##ost
+too
+##cial
+world
+give
+##ike
+Okay
+##ways
+min
+##ward
+##ily
+gen
+find
+dec
+##ular
+##ob
+tell
+Now
+sm
+cour
+real
+##cess
+##nds
+big
+num
+##ction
+add
+set
+um
+##ood
+##ible
+own
+life
+##ities
+its
+God
+##pect
+didn
+##stem
+##les
+##uc
+##ib
+##ating
+##olog
+person
+inv
+##ably
+sure
+reg
+##lic
+stu
+cr
+ev
+##ments
+another
+la
+last
+sub
+att
+op
+inst
+sl
+happen
+rep
+import
+##ific
+##ix
+made
+ear
+ac
+def
+##ute
+next
+##ative
+form
+guys
+system
+##ew
+able
+##ied
+always
+##ren
+##erest
+As
+mod
+done
+##ings
+love
+##ism
+ask
+##old
+##ered
+trans
+count
+##ility
+high
+fin
+prob
+pol
+exam
+pres
+maybe
+##ell
+stud
+prod
+car
+##ock
+used
+##oy
+##stand
+No
+mon
+##ks
+interest
+ent
+##ited
+sort
+For
+today
+##ics
+vide
+bec
+Well
+Al
+important
+such
+run
+keep
+fact
+##ata
+##ss
+never
+##ween
+stuff
+##ract
+question
+##als
+sim
+##vern
+##ather
+course
+Of
+##oc
+##ness
+##arch
+##ize
+All
+##ense
+##blem
+probably
+##hip
+number
+##ention
+saying
+commun
+An
+##akes
+belie
+between
+better
+##cus
+place
+gener
+ca
+ins
+ass
+##cond
+##cept
+##ull
+understand
+fun
+thought
+##gan
+##iew
+##cy
+##ution
+##ope
+##ason
+problem
+doesn
+##ational
+read
+trying
+sch
+el
+##ah
+##atter
+exper
+four
+ele
+cou
+##ont
+called
+partic
+open
+gl
+everything
+eff
+getting
+ty
+Am
+Because
+##ave
+met
+Like
+##oney
+▁
+##e
+##t
+##o
+##a
+##n
+##i
+##s
+##h
+##r
+##l
+##d
+##u
+##c
+##y
+##m
+##g
+##w
+##f
+##p
+##,
+##b
+##.
+##k
+##v
+##'
+##I
+##T
+##A
+##S
+##j
+##x
+##W
+##B
+##C
+##?
+##0
+##O
+##-
+##M
+##H
+##Y
+##q
+##1
+##P
+##z
+##L
+##D
+##N
+##G
+##F
+##R
+##E
+##2
+##J
+##U
+##:
+##5
+##9
+##3
+##K
+##4
+##V
+##8
+##6
+##7
+##!
+##%
+##Q
+##$
+##Z
+##X
+##é
+##/
+##í
+##á
+##£
+##ó
+##ā
+##ü
+##ñ
+##ö
+##è
+##ç
+##à
+##¿
+##μ
+##π
+##ä
+##ú
+##θ
+##ã
+##φ
+##ī
+##σ
+##ê
+##ρ
+##â
+##ô
+##^
+##€
+##É
+##ū
+##Δ
+##λ
+##α
+##τ
+##æ
+##а
+##о
+##ν
+##î
+##γ
+##ψ
+##ē
+##т
+##ß
+##ω
+##ï
+##ć
+##č
+##ε
+##е
+##и
+##ò
+##р
+##β
+##ø
+##ł
+##δ
+##η
+##п
+##ë
+##н
+##с
+##š
+##Ü
+##å
+##ń
+##ś
+##я
+##đ
+##л
+##м
+##Ö
+##û
+##ș
+##в
+##Á
+##Ø
+##ù
+##ο
+##ч
+##ь
+##ž
+##Φ
+##у
+##ę
+##ι
+##б
+##г
+##к
+##ő
+##Ś
+##Ω
+##κ
+##υ
+##ì
+##Č
+##έ
+##х
+##ы
+##Å
+##Ç
+##ż
+##ί
+##ζ
+##χ
+##э
+##Æ
+##Í
+##õ
+##ě
+##ħ
+##Ł
+##œ
+##Ž
+##ț
+##Γ
+##П
+##д
+##з
+##ф
+##¡
+##À
+##Î
+##Ā
+##ė
+##Š
+##ź
+##Κ
+##Ψ
+##ά
+##ξ
+##ό

ASR/parakeet/artifacts/705f11d22dc04b169effc35ce5cd1361_tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a3a82c48998709f9bc9f5db2924d99a899d20bf82550ca52272da7c7557b0a0
+size 251396

ASR/parakeet/artifacts/a4715c7f6b2d4c2bb709306073d0c0a4_tokenizer.vocab

@@ -0,0 +1,1024 @@
+<unk>	0
+▁t	-0
+▁th	-1
+▁a	-2
+in	-3
+▁the	-4
+re	-5
+▁w	-6
+▁o	-7
+▁s	-8
+at	-9
+ou	-10
+er	-11
+nd	-12
+▁i	-13
+▁b	-14
+▁c	-15
+on	-16
+▁h	-17
+ing	-18
+▁to	-19
+▁m	-20
+en	-21
+▁f	-22
+▁p	-23
+an	-24
+▁d	-25
+es	-26
+or	-27
+ll	-28
+▁of	-29
+▁and	-30
+▁y	-31
+▁l	-32
+▁I	-33
+it	-34
+▁in	-35
+is	-36
+ed	-37
+▁g	-38
+▁you	-39
+ar	-40
+▁that	-41
+om	-42
+as	-43
+▁n	-44
+ve	-45
+us	-46
+ic	-47
+ow	-48
+al	-49
+▁it	-50
+▁be	-51
+▁wh	-52
+le	-53
+ion	-54
+ut	-55
+ot	-56
+▁we	-57
+▁is	-58
+▁e	-59
+et	-60
+ay	-61
+▁re	-62
+▁on	-63
+▁T	-64
+▁A	-65
+▁ha	-66
+ent	-67
+ke	-68
+ct	-69
+▁S	-70
+ig	-71
+ver	-72
+▁Th	-73
+all	-74
+id	-75
+▁for	-76
+ro	-77
+▁he	-78
+se	-79
+▁this	-80
+ld	-81
+ly	-82
+▁go	-83
+▁k	-84
+▁st	-85
+st	-86
+ch	-87
+▁li	-88
+▁u	-89
+am	-90
+ur	-91
+ce	-92
+ith	-93
+im	-94
+▁so	-95
+▁have	-96
+▁do	-97
+ht	-98
+th	-99
+▁an	-100
+▁with	-101
+ad	-102
+▁r	-103
+ir	-104
+▁was	-105
+▁as	-106
+▁W	-107
+▁are	-108
+ust	-109
+ally	-110
+▁j	-111
+▁se	-112
+ation	-113
+od	-114
+ere	-115
+▁like	-116
+▁not	-117
+▁kn	-118
+ight	-119
+▁B	-120
+▁they	-121
+▁And	-122
+▁know	-123
+ome	-124
+op	-125
+▁can	-126
+▁or	-127
+▁sh	-128
+▁me	-129
+ill	-130
+ant	-131
+ck	-132
+▁what	-133
+▁at	-134
+▁ab	-135
+ould	-136
+ol	-137
+▁So	-138
+▁C	-139
+use	-140
+ter	-141
+il	-142
+▁but	-143
+▁just	-144
+▁ne	-145
+▁de	-146
+ra	-147
+ore	-148
+▁there	-149
+ul	-150
+out	-151
+▁con	-152
+▁all	-153
+▁The	-154
+ers	-155
+▁H	-156
+▁fr	-157
+▁pro	-158
+ge	-159
+ea	-160
+▁Y	-161
+▁O	-162
+▁M	-163
+pp	-164
+▁com	-165
+ess	-166
+▁ch	-167
+▁al	-168
+est	-169
+ate	-170
+qu	-171
+▁lo	-172
+▁ex	-173
+very	-174
+▁su	-175
+ain	-176
+▁one	-177
+ca	-178
+art	-179
+ist	-180
+if	-181
+ive	-182
+▁if	-183
+ink	-184
+nt	-185
+ab	-186
+▁about	-187
+▁going	-188
+▁v	-189
+▁wor	-190
+um	-191
+ok	-192
+▁your	-193
+▁my	-194
+ind	-195
+▁get	-196
+cause	-197
+▁from	-198
+▁don	-199
+ri	-200
+pe	-201
+un	-202
+ity	-203
+▁up	-204
+▁P	-205
+▁out	-206
+ort	-207
+▁L	-208
+ment	-209
+el	-210
+▁N	-211
+▁some	-212
+ich	-213
+and	-214
+▁think	-215
+em	-216
+oug	-217
+▁G	-218
+os	-219
+▁D	-220
+res	-221
+▁because	-222
+▁by	-223
+ake	-224
+▁int	-225
+ie	-226
+▁us	-227
+▁tr	-228
+▁then	-229
+ack	-230
+▁pl	-231
+▁here	-232
+▁pe	-233
+her	-234
+▁will	-235
+▁F	-236
+▁which	-237
+ard	-238
+▁right	-239
+▁thing	-240
+▁want	-241
+ies	-242
+ople	-243
+▁It	-244
+▁them	-245
+ame	-246
+▁We	-247
+our	-248
+▁say	-249
+▁R	-250
+▁people	-251
+▁see	-252
+▁who	-253
+ast	-254
+ure	-255
+ect	-256
+ear	-257
+▁tim	-258
+▁E	-259
+▁You	-260
+▁would	-261
+▁when	-262
+ven	-263
+▁our	-264
+ci	-265
+▁really	-266
+▁more	-267
+ound	-268
+ose	-269
+ak	-270
+▁co	-271
+ide	-272
+ough	-273
+▁had	-274
+so	-275
+▁qu	-276
+eah	-277
+▁were	-278
+ine	-279
+▁act	-280
+ther	-281
+▁these	-282
+▁how	-283
+▁now	-284
+▁sa	-285
+ud	-286
+▁Wh	-287
+▁man	-288
+ous	-289
+one	-290
+pt	-291
+ff	-292
+ong	-293
+▁has	-294
+▁any	-295
+▁very	-296
+▁But	-297
+▁look	-298
+iv	-299
+itt	-300
+▁time	-301
+▁mo	-302
+▁ar	-303
+hing	-304
+▁le	-305
+▁work	-306
+▁their	-307
+are	-308
+▁his	-309
+per	-310
+ions	-311
+▁im	-312
+▁ag	-313
+▁J	-314
+▁no	-315
+▁en	-316
+▁got	-317
+ag	-318
+▁sp	-319
+ans	-320
+act	-321
+▁te	-322
+▁also	-323
+iz	-324
+ice	-325
+▁That	-326
+▁cl	-327
+▁been	-328
+▁way	-329
+▁fe	-330
+▁did	-331
+ple	-332
+ually	-333
+▁other	-334
+▁U	-335
+ite	-336
+age	-337
+omet	-338
+ber	-339
+reat	-340
+ree	-341
+▁into	-342
+own	-343
+▁tw	-344
+▁part	-345
+alk	-346
+▁where	-347
+▁need	-348
+▁every	-349
+pl	-350
+▁ad	-351
+ry	-352
+▁over	-353
+ble	-354
+ap	-355
+ue	-356
+▁kind	-357
+▁po	-358
+▁back	-359
+▁cont	-360
+iff	-361
+▁somet	-362
+▁pr	-363
+nder	-364
+ire	-365
+▁good	-366
+▁than	-367
+ace	-368
+▁gu	-369
+ep	-370
+og	-371
+ick	-372
+way	-373
+▁lot	-374
+▁un	-375
+▁things	-376
+▁In	-377
+ish	-378
+kay	-379
+▁well	-380
+▁could	-381
+▁pre	-382
+▁two	-383
+irst	-384
+▁diff	-385
+ach	-386
+cc	-387
+ittle	-388
+int	-389
+▁He	-390
+▁those	-391
+ence	-392
+ip	-393
+ase	-394
+▁him	-395
+▁make	-396
+▁little	-397
+ical	-398
+▁gr	-399
+▁year	-400
+ass	-401
+▁thr	-402
+uch	-403
+ated	-404
+▁This	-405
+▁off	-406
+▁res	-407
+ac	-408
+ance	-409
+▁actually	-410
+▁talk	-411
+ult	-412
+able	-413
+orm	-414
+▁dis	-415
+▁first	-416
+ations	-417
+▁something	-418
+▁she	-419
+sel	-420
+▁let	-421
+ord	-422
+▁may	-423
+ia	-424
+▁am	-425
+▁her	-426
+▁said	-427
+▁bo	-428
+be	-429
+ount	-430
+▁much	-431
+▁per	-432
+▁even	-433
+▁differe	-434
+vel	-435
+ary	-436
+▁app	-437
+ving	-438
+▁comm	-439
+▁imp	-440
+ys	-441
+▁again	-442
+ress	-443
+▁yeah	-444
+▁down	-445
+ang	-446
+▁mean	-447
+na	-448
+ens	-449
+▁does	-450
+▁fo	-451
+▁comp	-452
+▁ro	-453
+▁bl	-454
+ody	-455
+▁K	-456
+▁through	-457
+▁start	-458
+uct	-459
+▁only	-460
+▁bet	-461
+▁under	-462
+▁br	-463
+▁take	-464
+ning	-465
+▁bu	-466
+▁use	-467
+▁Ch	-468
+xt	-469
+co	-470
+ory	-471
+ild	-472
+▁put	-473
+▁call	-474
+▁new	-475
+other	-476
+ting	-477
+▁happ	-478
+ater	-479
+▁inc	-480
+ition	-481
+▁different	-482
+▁should	-483
+ade	-484
+ign	-485
+thing	-486
+▁day	-487
+fore	-488
+▁Yeah	-489
+ark	-490
+ile	-491
+ial	-492
+▁come	-493
+▁They	-494
+▁being	-495
+▁try	-496
+ious	-497
+▁sc	-498
+▁bit	-499
+▁spe	-500
+ub	-501
+fe	-502
+▁doing	-503
+▁St	-504
+vers	-505
+av	-506
+ty	-507
+ian	-508
+onna	-509
+red	-510
+wn	-511
+▁ke	-512
+form	-513
+ors	-514
+▁fl	-515
+fter	-516
+ail	-517
+ents	-518
+▁gonna	-519
+▁point	-520
+ces	-521
+▁There	-522
+self	-523
+▁many	-524
+▁If	-525
+▁same	-526
+▁sy	-527
+▁quest	-528
+▁most	-529
+▁great	-530
+▁What	-531
+▁fu	-532
+ug	-533
+▁show	-534
+we	-535
+ual	-536
+ons	-537
+▁Be	-538
+ically	-539
+▁ser	-540
+▁rem	-541
+▁ind	-542
+▁pers	-543
+▁V	-544
+he	-545
+▁str	-546
+ved	-547
+▁still	-548
+ank	-549
+▁rec	-550
+▁wr	-551
+ought	-552
+day	-553
+ath	-554
+▁end	-555
+▁bas	-556
+ft	-557
+erm	-558
+body	-559
+ph	-560
+ject	-561
+ict	-562
+▁play	-563
+▁Is	-564
+ates	-565
+▁ph	-566
+oth	-567
+▁acc	-568
+get	-569
+▁years	-570
+▁em	-571
+▁id	-572
+▁Oh	-573
+ves	-574
+ever	-575
+▁inter	-576
+▁rel	-577
+▁before	-578
+▁feel	-579
+igh	-580
+▁three	-581
+iss	-582
+▁des	-583
+ne	-584
+▁why	-585
+▁uh	-586
+▁To	-587
+▁cons	-588
+▁hel	-589
+▁after	-590
+ower	-591
+urn	-592
+▁okay	-593
+▁long	-594
+▁bel	-595
+▁around	-596
+ful	-597
+te	-598
+ise	-599
+▁ob	-600
+▁supp	-601
+ady	-602
+ange	-603
+aking	-604
+▁pos	-605
+atch	-606
+▁tra	-607
+gr	-608
+▁might	-609
+ert	-610
+▁help	-611
+ost	-612
+▁too	-613
+cial	-614
+▁world	-615
+▁give	-616
+ike	-617
+▁Okay	-618
+ways	-619
+▁min	-620
+ward	-621
+ily	-622
+▁gen	-623
+▁find	-624
+▁dec	-625
+ular	-626
+ob	-627
+▁tell	-628
+▁Now	-629
+▁sm	-630
+▁cour	-631
+▁real	-632
+cess	-633
+nds	-634
+▁big	-635
+▁num	-636
+ction	-637
+▁add	-638
+▁set	-639
+▁um	-640
+ood	-641
+ible	-642
+▁own	-643
+▁life	-644
+ities	-645
+▁its	-646
+▁God	-647
+pect	-648
+▁didn	-649
+stem	-650
+les	-651
+uc	-652
+ib	-653
+ating	-654
+olog	-655
+▁person	-656
+▁inv	-657
+ably	-658
+▁sure	-659
+▁reg	-660
+lic	-661
+▁stu	-662
+▁cr	-663
+▁ev	-664
+ments	-665
+▁another	-666
+▁la	-667
+▁last	-668
+▁sub	-669
+▁att	-670
+▁op	-671
+▁inst	-672
+▁sl	-673
+▁happen	-674
+▁rep	-675
+▁import	-676
+ific	-677
+ix	-678
+▁made	-679
+▁ear	-680
+▁ac	-681
+▁def	-682
+ute	-683
+▁next	-684
+ative	-685
+▁form	-686
+▁guys	-687
+▁system	-688
+ew	-689
+▁able	-690
+ied	-691
+▁always	-692
+ren	-693
+erest	-694
+▁As	-695
+▁mod	-696
+▁done	-697
+ings	-698
+▁love	-699
+ism	-700
+▁ask	-701
+old	-702
+ered	-703
+▁trans	-704
+▁count	-705
+ility	-706
+▁high	-707
+▁fin	-708
+▁prob	-709
+▁pol	-710
+▁exam	-711
+▁pres	-712
+▁maybe	-713
+ell	-714
+▁stud	-715
+▁prod	-716
+▁car	-717
+ock	-718
+▁used	-719
+oy	-720
+stand	-721
+▁No	-722
+▁mon	-723
+ks	-724
+▁interest	-725
+▁ent	-726
+ited	-727
+▁sort	-728
+▁For	-729
+▁today	-730
+ics	-731
+▁vide	-732
+▁bec	-733
+▁Well	-734
+▁Al	-735
+▁important	-736
+▁such	-737
+▁run	-738
+▁keep	-739
+▁fact	-740
+ata	-741
+ss	-742
+▁never	-743
+ween	-744
+▁stuff	-745
+ract	-746
+▁question	-747
+als	-748
+▁sim	-749
+vern	-750
+ather	-751
+▁course	-752
+▁Of	-753
+oc	-754
+ness	-755
+arch	-756
+ize	-757
+▁All	-758
+ense	-759
+blem	-760
+▁probably	-761
+hip	-762
+▁number	-763
+ention	-764
+▁saying	-765
+▁commun	-766
+▁An	-767
+akes	-768
+▁belie	-769
+▁between	-770
+▁better	-771
+cus	-772
+▁place	-773
+▁gener	-774
+▁ca	-775
+▁ins	-776
+▁ass	-777
+cond	-778
+cept	-779
+ull	-780
+▁understand	-781
+▁fun	-782
+▁thought	-783
+gan	-784
+iew	-785
+cy	-786
+ution	-787
+ope	-788
+ason	-789
+▁problem	-790
+▁doesn	-791
+ational	-792
+▁read	-793
+▁trying	-794
+▁sch	-795
+▁el	-796
+ah	-797
+atter	-798
+▁exper	-799
+▁four	-800
+▁ele	-801
+▁cou	-802
+ont	-803
+▁called	-804
+▁partic	-805
+▁open	-806
+▁gl	-807
+▁everything	-808
+▁eff	-809
+▁getting	-810
+▁ty	-811
+▁Am	-812
+▁Because	-813
+ave	-814
+▁met	-815
+▁Like	-816
+oney	-817
+▁	-818
+e	-819
+t	-820
+o	-821
+a	-822
+n	-823
+i	-824
+s	-825
+h	-826
+r	-827
+l	-828
+d	-829
+u	-830
+c	-831
+y	-832
+m	-833
+g	-834
+w	-835
+f	-836
+p	-837
+,	-838
+b	-839
+.	-840
+k	-841
+v	-842
+'	-843
+I	-844
+T	-845
+A	-846
+S	-847
+j	-848
+x	-849
+W	-850
+B	-851
+C	-852
+?	-853
+0	-854
+O	-855
+-	-856
+M	-857
+H	-858
+Y	-859
+q	-860
+1	-861
+P	-862
+z	-863
+L	-864
+D	-865
+N	-866
+G	-867
+F	-868
+R	-869
+E	-870
+2	-871
+J	-872
+U	-873
+:	-874
+5	-875
+9	-876
+3	-877
+K	-878
+4	-879
+V	-880
+8	-881
+6	-882
+7	-883
+!	-884
+%	-885
+Q	-886
+$	-887
+Z	-888
+X	-889
+é	-890
+/	-891
+í	-892
+á	-893
+£	-894
+ó	-895
+ā	-896
+ü	-897
+ñ	-898
+ö	-899
+è	-900
+ç	-901
+à	-902
+¿	-903
+μ	-904
+π	-905
+ä	-906
+ú	-907
+θ	-908
+ã	-909
+φ	-910
+ī	-911
+σ	-912
+ê	-913
+ρ	-914
+â	-915
+ô	-916
+^	-917
+€	-918
+É	-919
+ū	-920
+Δ	-921
+λ	-922
+α	-923
+τ	-924
+æ	-925
+а	-926
+о	-927
+ν	-928
+î	-929
+γ	-930
+ψ	-931
+ē	-932
+т	-933
+ß	-934
+ω	-935
+ï	-936
+ć	-937
+č	-938
+ε	-939
+е	-940
+и	-941
+ò	-942
+р	-943
+β	-944
+ø	-945
+ł	-946
+δ	-947
+η	-948
+п	-949
+ë	-950
+н	-951
+с	-952
+š	-953
+Ü	-954
+å	-955
+ń	-956
+ś	-957
+я	-958
+đ	-959
+л	-960
+м	-961
+Ö	-962
+û	-963
+ș	-964
+в	-965
+Á	-966
+Ø	-967
+ù	-968
+ο	-969
+ч	-970
+ь	-971
+ž	-972
+Φ	-973
+у	-974
+ę	-975
+ι	-976
+б	-977
+г	-978
+к	-979
+ő	-980
+Ś	-981
+Ω	-982
+κ	-983
+υ	-984
+ì	-985
+Č	-986
+έ	-987
+х	-988
+ы	-989
+Å	-990
+Ç	-991
+ż	-992
+ί	-993
+ζ	-994
+χ	-995
+э	-996
+Æ	-997
+Í	-998
+õ	-999
+ě	-1000
+ħ	-1001
+Ł	-1002
+œ	-1003
+Ž	-1004
+ț	-1005
+Γ	-1006
+П	-1007
+д	-1008
+з	-1009
+ф	-1010
+¡	-1011
+À	-1012
+Î	-1013
+Ā	-1014
+ė	-1015
+Š	-1016
+ź	-1017
+Κ	-1018
+Ψ	-1019
+ά	-1020
+ξ	-1021
+ό	-1022

ASR/parakeet/model.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f2ff18de72abdbe7fbde1dc50be1cf5dd156f58d258940c25d33db14cc23bfc
+size 7415423311

ASR/{README.md → whisper/README.md}

@@ -26,10 +26,12 @@ model-index:
             value: 0.66
 ---
 
-# Whisper Large v3 — Singapore Military ATC (CTranslate2 float16)
+# Whisper Large v3 - Singapore Military ATC (Legacy CTranslate2 Backend)
 
 Fine-tuned Whisper Large v3 for Singapore Air Force air traffic control speech recognition.
 
+> **Legacy backend.** This folder is kept for compatibility with the original ASTRA faster-whisper / CTranslate2 inference path. The newer NeMo checkpoint lives in [../parakeet](../parakeet).
+
 ## Performance
 
 | Run | WER | Base | Data | Key Change |
@@ -95,7 +97,7 @@ See [hyperparameters.md](./hyperparameters.md) for full training configuration.
 ```python
 from faster_whisper import WhisperModel
 
-model = WhisperModel("path/to/ASR", device="cuda", compute_type="float16")
+model = WhisperModel("path/to/ASR/whisper", device="cuda", compute_type="float16")
 segments, info = model.transcribe(
     "audio.wav",
     language="en",
@@ -113,7 +115,6 @@ segments, info = model.transcribe(
     ),
 )
 text = " ".join(seg.text.strip() for seg in segments)
-# "camel cleared i l s approach runway three six"
 ```
 
 ## Output Format
@@ -126,4 +127,4 @@ The model outputs **normalized spoken text** (lowercase, fully expanded):
 | "Contact Tengah Approach one three zero decimal zero" | `contact tengah approach one three zero decimal zero` |
 | "Squawk seven seven zero zero" | `squawk seven seven zero zero` |
 
-A companion rule-based formatter (23 deterministic rules, <1ms, 0 VRAM) converts to display text (e.g., `CAMEL climb FL090`). See the [ASTRA simpilot](https://github.com/aether-raid) pipeline for the full integration.
+A companion rule-based formatter converts this into display text such as `CAMEL climb FL090`.

README.md

@@ -3,6 +3,8 @@ language:
   - en
 license: other
 tags:
+  - nemo
+  - parakeet
   - whisper
   - qwen3
   - ctranslate2
@@ -21,30 +23,41 @@ Fine-tuned models for Singapore military air traffic control, built for the [AST
 
 ## Pipeline
 
-```
-Audio  -->  VAD (Silero)  -->  ASR (Whisper)  -->  Rule Formatter  -->  Display Text
+```text
+Audio  -->  VAD (Silero)  -->  ASR (Whisper or Parakeet)  -->  Rule Formatter  -->  Display Text
                                "camel climb flight level zero nine zero"
-                                                                        "CAMEL climb FL090"
+                                                                                  "CAMEL climb FL090"
 ```
 
-The production pipeline uses a **rule-based formatter** (23 deterministic rules, <1ms, 0 VRAM) instead of the LLM. The LLM is retained for reference.
+The production pipeline uses a deterministic rule-based formatter instead of the legacy LLM formatter.
 
 ## Models
 
-### [ASR/](./ASR) — Whisper Large v3 (CTranslate2 float16)
+### [ASR/whisper/](./ASR/whisper) - Whisper Large v3 (Legacy CTranslate2 backend)
 
 Fine-tuned for Singapore military ATC speech. Uses CTranslate2 float16 format for fast inference with [faster-whisper](https://github.com/SYSTRAN/faster-whisper).
 
 | Metric | Value |
 |--------|-------|
-| WER | **0.82%** |
+| WER | **0.66%** |
 | Base model | `openai/whisper-large-v3` |
 | Size | 2.9 GB |
-| Training | Full fine-tune with enhanced VHF radio augmentation |
+| Runtime | `faster-whisper` / CTranslate2 |
+
+### [ASR/parakeet/](./ASR/parakeet) - Parakeet-TDT 0.6B v2 (NeMo checkpoint)
+
+Fine-tuned NeMo Parakeet model for Singapore military ATC speech. Published as a raw checkpoint together with the tokenizer artifacts required to restore it.
+
+| Metric | Value |
+|--------|-------|
+| Validation WER | **1.14%** |
+| Base model | `nvidia/parakeet-tdt-0.6b-v2` |
+| Size | 7.0 GB |
+| Runtime | `nemo_toolkit[asr]` |
 
-### [LLM/](./LLM) — Qwen3-1.7B Display Formatter (Legacy)
+### [LLM/](./LLM) - Qwen3-1.7B Display Formatter (Legacy)
 
-> **Legacy.** Superseded by a deterministic rule-based formatter. Retained for reference.
+> **Legacy.** Superseded by the deterministic rule formatter. Retained for reference only.
 
 Converts normalized ASR output into structured ATC display text.
 
@@ -56,26 +69,20 @@ Converts normalized ASR output into structured ATC display text.
 
 ## Architecture
 
+```text
+Audio --> VAD (Silero) --> ASR backend --> Post-processing --> Rule Formatter --> Display Text
 ```
-Audio --> VAD (Silero) --> ASR (Whisper ct2) --> Post-processing --> Rule Formatter --> Display Text
-```
-
-| Component | Technology | Latency | VRAM |
-|-----------|-----------|---------|------|
-| VAD | Silero VAD (ONNX) | ~50ms | <100 MB |
-| ASR | Whisper Large v3 (CTranslate2) | ~500ms-2s | ~2 GB |
-| Formatter | 23 deterministic rules | <1ms | 0 MB |
 
-Total VRAM: ~2 GB (ASR only).
+| Component | Technology | Notes |
+|-----------|------------|-------|
+| VAD | Silero VAD | Shared frontend for both ASR backends |
+| ASR (legacy) | Whisper Large v3 (CTranslate2) | Lower-memory legacy backend |
+| ASR (current NeMo path) | Parakeet-TDT 0.6B v2 | Fine-tuned NeMo checkpoint |
+| Formatter | Deterministic rules | Converts normalized speech to ATC display text |
 
 ## Domain
 
-Singapore military ATC covering:
-- **Airbases**: Tengah (WSAT, runway 18/36), Paya Lebar (WSAP, runway 02/20)
-- **Aircraft**: F-16C/D, F-15SG, C-130, Hercules
-- **Approaches**: ILS, GCA, PAR, TACAN, DVOR/DME, VOR/DME, Visual Straight-in
-- **100+ callsigns**: CAMEL, NINJA, BEETLE, TAIPAN, MAVERICK, JAGUAR, LANCER, etc.
-- **Categories**: departure, approach, handoff, maneuver, landing, emergency, ground, recovery, pilot reports, military-specific ops
+Singapore military ATC covering Tengah and Paya Lebar operations, military phraseology, 100+ callsigns, and approach / recovery / emergency traffic.
 
 ## Training History
 
@@ -87,9 +94,9 @@ Singapore military ATC covering:
 | ct2_run6 | 0.40% | jacktol/whisper-large-v3-finetuned-for-ATC | +augmentation, weight decay |
 | ct2_run7 | 0.24% | jacktol/whisper-large-v3-finetuned-for-ATC | Frozen encoder, +50 real recordings |
 | ct2_run8 | 0.66% | openai/whisper-large-v3 | Full retrain from base, enhanced augmentation |
-| **ct2_run9** | **0.82%** | openai/whisper-large-v3 | Expanded dataset (+MNSC, +deepdml, 17.8k train), 19 epochs |
+| parakeet_atc | 1.14% | nvidia/parakeet-tdt-0.6b-v2 | NeMo fine-tune, best checkpoint at epoch 20 |
 
-### LLM (Legacy)
+### LLM
 
 | Run | Accuracy | Key Change |
 |-----|----------|------------|
@@ -98,24 +105,31 @@ Singapore military ATC covering:
 
 ## Quick Start
 
-### ASR
+### Whisper ASR
 
 ```python
 from faster_whisper import WhisperModel
 
-model = WhisperModel("./ASR", device="cuda", compute_type="float16")
+model = WhisperModel("./ASR/whisper", device="cuda", compute_type="float16")
 segments, info = model.transcribe("audio.wav", language="en", beam_size=5)
 text = " ".join(seg.text.strip() for seg in segments)
 ```
 
-### Download
+### Parakeet ASR
+
+See [ASR/parakeet/README.md](./ASR/parakeet/README.md) for the NeMo restore example and tokenizer artifact requirements.
+
+## Download
 
 ```bash
-# Full repo (ASR + LLM)
+# Full repo
 huggingface-cli download aether-raid/astra-atc-models --local-dir ./models
 
-# ASR only (recommended)
-huggingface-cli download aether-raid/astra-atc-models --include "ASR/*" --local-dir ./models
+# Whisper ASR only
+huggingface-cli download aether-raid/astra-atc-models --include "ASR/whisper/*" --local-dir ./models
+
+# Parakeet ASR only
+huggingface-cli download aether-raid/astra-atc-models --include "ASR/parakeet/*" --local-dir ./models
 
 # LLM only (legacy)
 huggingface-cli download aether-raid/astra-atc-models --include "LLM/*" --local-dir ./models