Datasets:
Mise à jour de la carte du dataset
Browse files- README.md +163 -235
- media/noise_chain.mp4 +3 -0
README.md
CHANGED
|
@@ -32,260 +32,188 @@ configs:
|
|
| 32 |
path: data/STEAK-tough/test-*.parquet
|
| 33 |
---
|
| 34 |
|
| 35 |
-
# STEAK — **S**peech-to-**T**ext **E**
|
| 36 |
-
|
| 37 |
-
ATC synthétique. **Une ligne = un audio** (un énoncé contrôleur ATCO *ou* un
|
| 38 |
-
readback pilote). 2,519,694 audios.
|
| 39 |
-
La **paire ATCO↔pilote est toujours présente** dans le dataset : les deux rôles
|
| 40 |
-
d'un échange partagent le même `pair_id` — filtrer/regrouper sur `pair_id` pour
|
| 41 |
-
reconstituer les paires complètes (voir « Regrouper les audios en paires »).
|
| 42 |
-
|
| 43 |
-
## Génération du texte (règles formelles + ontologie ATC)
|
| 44 |
-
|
| 45 |
-
Le texte de chaque échange n'est pas tiré d'enregistrements réels : il est
|
| 46 |
-
**généré par règles formelles** suivant une **ontologie de la communication
|
| 47 |
-
ATC**. Chaque énoncé du contrôleur est une **chaîne de commandes** assemblée
|
| 48 |
-
ainsi :
|
| 49 |
-
|
| 50 |
-
- **82 commandes** au total sont générables (table complète plus bas).
|
| 51 |
-
Un énoncé en contient **au plus 7** (1 commande *seed* + 6 ajoutées au maximum)
|
| 52 |
-
et **2,2 en moyenne** (1 seed + ~1,2 ajoutée, cf. tirage ci-dessous).
|
| 53 |
-
- La **première commande (seed)** est tirée au hasard parmi les 82
|
| 54 |
-
selon une **pondération** (toutes n'ont pas la même probabilité) : les poids
|
| 55 |
-
sont issus des **observations EDA** sur les corpus ATC réels **ATCO2** et
|
| 56 |
-
**UWB-ATCC**, pour que le cœur en-route (climb/descend/heading/contact…)
|
| 57 |
-
domine et que les commandes rares restent rares.
|
| 58 |
-
- **Ajout des commandes suivantes** : 6 slots, chacun rempli par un tirage de
|
| 59 |
-
**Bernoulli indépendant de probabilité 0,2** (≈ Binomiale(6 ; 0,2)). En
|
| 60 |
-
moyenne ~1,2 commande supplémentaire, parfois 0, parfois jusqu'à 6.
|
| 61 |
-
- **Règles d'incompatibilité** dans un même énoncé : au plus **une commande par
|
| 62 |
-
groupe** (jamais deux commandes verticales, ni deux commandes de cap, etc.),
|
| 63 |
-
cohérence du sens vertical (pas de « climb … rate of descent »), exclusivité
|
| 64 |
-
sol/vol (pas de « taxi … descend FL90 ») et conflits de phase
|
| 65 |
-
(départ ⊥ arrivée…).
|
| 66 |
-
- **Formulations multiples** : chaque commande dispose de plusieurs formulations
|
| 67 |
-
côté contrôleur, et de plusieurs façons d'être **relue** côté pilote — la
|
| 68 |
-
formulation du readback n'est **pas forcément la même** que celle de
|
| 69 |
-
l'instruction (parfois réduite à la valeur seule, « level 100 »).
|
| 70 |
-
- **Valeur** de chaque commande **tirée au hasard dans son domaine** de valeurs
|
| 71 |
-
possibles (cf. table ci-dessous).
|
| 72 |
-
|
| 73 |
-
Les valeurs concrètes (waypoints, noms et fréquences de stations, aéroports,
|
| 74 |
-
compagnies/indicatifs) proviennent de **sources publiques** :
|
| 75 |
-
|
| 76 |
-
| ressource | source |
|
| 77 |
-
|---|---|
|
| 78 |
-
| waypoints | [OpenNav](https://opennav.com/waypoint/FR) (une page par pays) |
|
| 79 |
-
| aéroports ICAO | [OpenNav](https://opennav.com/airportcodes) |
|
| 80 |
-
| stations ATC + fréquences | [VATEUD](https://fsmine.dhis.org/vateud8/) |
|
| 81 |
-
| compagnies / indicatifs (telephony) | [Wikipedia](https://en.wikipedia.org/wiki/List_of_airline_codes), [flugzeuginfo.net](https://www.flugzeuginfo.net/), [OpenNav](https://opennav.com/airlinecodes/), [FAA CNT](https://www.faa.gov/air_traffic/publications/atpubs/cnt_html/) |
|
| 82 |
-
| voix de référence seed-vc (accent, `accent_ref`) | [Speech Accent Archive](https://accent.gmu.edu) (≈ 3 000 locuteurs de toutes langues maternelles → accents variés ; cf. « Rendu audio ») |
|
| 83 |
-
|
| 84 |
-
### Conditions accrochées aux commandes
|
| 85 |
-
|
| 86 |
-
Une commande peut recevoir une **condition** (déclencheur temporel/spatial). Le
|
| 87 |
-
mécanisme :
|
| 88 |
-
|
| 89 |
-
- Avec une probabilité de **8 %**, l'énoncé reçoit **une** condition
|
| 90 |
-
(dans 92 % des cas, aucune).
|
| 91 |
-
- Une **commande hôte** est tirée parmi celles de l'énoncé éligibles, et la
|
| 92 |
-
condition est filtrée sur **le groupe de cette hôte** (une condition verticale
|
| 93 |
-
ne s'accroche qu'à une commande verticale, etc.).
|
| 94 |
-
- Le connecteur est tiré (pondéré) parmi ceux applicables à ce groupe ; certains
|
| 95 |
-
portent une **valeur** (waypoint, FL, altitude, distance, vario, heure) tirée
|
| 96 |
-
dans son domaine — un waypoint de condition ne réutilise jamais un waypoint
|
| 97 |
-
déjà cité dans l'énoncé.
|
| 98 |
-
- Avec une probabilité de **18 %**, une **2e** condition (à
|
| 99 |
-
connecteur différent) est chaînée à la même hôte (ex. « when established …
|
| 100 |
-
until 4 miles final »).
|
| 101 |
-
- Côté pilote, la condition est relue avec une probabilité de
|
| 102 |
-
**55 %** (pas toujours relue, comme sur les vraies
|
| 103 |
-
fréquences).
|
| 104 |
-
|
| 105 |
-
La condition apparaît dans la colonne `commandes` (`atco_condition`) et, si relue,
|
| 106 |
-
dans `pilot_condition_readback`.
|
| 107 |
|
| 108 |
-
|
| 109 |
-
|
| 110 |
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
| « when established » | approche, verticales (climb / descend / maintain…) |
|
| 116 |
-
| « when passing … feet » | verticales (climb / descend / maintain…) |
|
| 117 |
-
| « until further advised » | cap, verticales (climb / descend / maintain…), vitesse |
|
| 118 |
-
| « until passing flight level … » | verticales (climb / descend / maintain…) |
|
| 119 |
-
| « until … » | direct-to / inbound, verticales (climb / descend / maintain…) |
|
| 120 |
-
| « until … miles » | verticales (climb / descend / maintain…), vitesse |
|
| 121 |
-
| « after … » | cap, direct-to / inbound, verticales (climb / descend / maintain…) |
|
| 122 |
-
| « after … then … » | cap, direct-to / inbound, verticales (climb / descend / maintain…) |
|
| 123 |
-
| « after the low approach » | cap, verticales (climb / descend / maintain…) |
|
| 124 |
-
| « after departure » | direct-to / inbound, verticales (climb / descend / maintain…) |
|
| 125 |
-
| « level by … » | verticales (climb / descend / maintain…) |
|
| 126 |
-
| « before … » | direct-to / inbound, verticales (climb / descend / maintain…) |
|
| 127 |
-
| « at … feet per minute » | verticales (climb / descend / maintain…) |
|
| 128 |
-
| « at … feet per minute or greater » | verticales (climb / descend / maintain…) |
|
| 129 |
-
| « time … » | cross waypoint, direct-to / inbound, verticales (climb / descend / maintain…) |
|
| 130 |
-
| « from present » | cap, direct-to / inbound, verticales (climb / descend / maintain…) |
|
| 131 |
-
| « from present position » | cap, direct-to / inbound |
|
| 132 |
-
| « if able » | direct-to / inbound, verticales (climb / descend / maintain…), vitesse |
|
| 133 |
-
| « if possible » | direct-to / inbound, verticales (climb / descend / maintain…), vitesse |
|
| 134 |
-
| « if available » | verticales (climb / descend / maintain…) |
|
| 135 |
-
| « until … miles final » | approche, verticales (climb / descend / maintain…), vitesse |
|
| 136 |
-
| « on conversion » | vitesse |
|
| 137 |
-
| « when established on the localizer » | approche, verticales (climb / descend / maintain…), vitesse |
|
| 138 |
-
| « when you are done de icing » | contact fréquence |
|
| 139 |
|
| 140 |
-
|
|
|
|
|
|
|
| 141 |
|
| 142 |
-
|
|
|
|
|
|
|
| 143 |
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
|
|
|
| 147 |
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
| Approche & atterrissage (cleared to land, cleared ILS / VOR / NDB / visual approach, touch-and-go, go-around…) | **piste 01–36** (+ suffixe **L / R / C**) |
|
| 152 |
-
| Navigation par waypoints (direct to, inbound, cross waypoint, follow route / SID / STAR / transition) | **waypoint** (source OpenNav) — éventuellement + FL ; routes **SID / STAR / transition** |
|
| 153 |
-
| Mouvements sur piste au sol (cross runway, hold short, line up, backtrack, vacate) | **piste 01–36** (+ suffixe **L / R / C**) ; taxiway pour *vacate* |
|
| 154 |
-
| Cap (turn left/right, turn by N°, heading) | cap **005–360°** (multiples de 5) ; sens **left / right** ; virage de **{10, 20, 30, 40, 90, 180, 270, 360}°** |
|
| 155 |
-
| QNH | **975–1045 hPa** |
|
| 156 |
-
| Transpondeur (squawk, ident) | code **0000–7777 octal** (codes d'urgence 7500/7600/7700/7777/0000 exclus) ; *ident* sans valeur |
|
| 157 |
-
| Vitesse (speed, reduce / increase speed, maintain speed, mach) | **160–340 kt** (pas de 10), Δ **10–50 kt**, ou **Mach 0.70–0.86** (pas de 0.01) |
|
| 158 |
-
| Transfert fréquence & report (contact, monitor, report passing) | **station + fréquence 118.000–136.xxx MHz** (canaux 8.33/25 kHz, source VATEUD) ; waypoint pour *report passing* |
|
| 159 |
-
| Roulage & départ sol (startup, pushback, taxi to / via, hold position) | identifiant de **taxiway** (sans valeur pour *startup / pushback / hold position*) |
|
| 160 |
-
| Décollage (cleared for take-off) | **piste 01–36** (+ suffixe **L / R / C**) |
|
| 161 |
-
| Clairances (IFR, destination, entrée CTR) | **aéroport ICAO** destination (source OpenNav) ; waypoint d'entrée CTR |
|
| 162 |
-
| Contrainte horaire (time) | **heure hhmm** (+ waypoint) |
|
| 163 |
-
| Attente (holding) | **waypoint** (source OpenNav) |
|
| 164 |
-
| Informations (ATIS, piste active, trafic, vent) | ATIS **alpha…zulu** ; piste active **01–36** ; trafic en position relative (« x o'clock y miles »…) ; vent **direction / force / rafale** |
|
| 165 |
-
| Réponses & initiation (affirm, standby, go ahead, appel initial) | — (sans valeur) |
|
| 166 |
-
|
| 167 |
-
## Rendu audio : voix, accent, bruit
|
| 168 |
-
|
| 169 |
-
Le texte est ensuite synthétisé puis dégradé pour ressembler à une vraie
|
| 170 |
-
fréquence ATC :
|
| 171 |
-
|
| 172 |
-
- **Voix / timbre** : chaque énoncé est synthétisé par **Kokoro TTS**, avec une
|
| 173 |
-
**voix tirée au hasard** parmi les voix Kokoro (colonne `voice`) — timbre varié
|
| 174 |
-
gratuitement.
|
| 175 |
-
- **Accent (conversion de voix)** : un étage **seed-vc** transfère ensuite la
|
| 176 |
-
voix d'un **clip de référence** sur la synthèse (foreign-accent conversion).
|
| 177 |
-
Les clips de référence (colonne `accent_ref`) sont **scrappés du
|
| 178 |
-
[Speech Accent Archive](https://accent.gmu.edu)** (≈ 3 000 locuteurs de toutes
|
| 179 |
-
langues maternelles lisant le même paragraphe → accents variés). Le pool est
|
| 180 |
-
réparti par genre et la part de voix masculines est tirée par rôle
|
| 181 |
-
(≈ 0,7 ATCO / ≈ 0,9 pilote).
|
| 182 |
-
- **Bruit** : un profil de bruit issu d'**enregistrements ATCO2 réels** est
|
| 183 |
-
ajouté avec un SNR tiré au hasard (cf. « Difficulté » et la colonne
|
| 184 |
-
`atco2_noise_profile`). Pas de réverbération (cabines ATC petites et
|
| 185 |
-
absorbantes → T60 négligeable).
|
| 186 |
-
|
| 187 |
-
## Difficulté (tiercile de RMS mesuré)
|
| 188 |
-
|
| 189 |
-
Difficulté = tiercile du **RMS mesuré sur le WAV rendu** (colonne `rms`).
|
| 190 |
-
|
| 191 |
-
Le **RMS** (*root-mean-square*) est le **niveau efficace** du signal : la
|
| 192 |
-
moyenne quadratique des échantillons, c.-à-d. l'énergie/le volume moyen de
|
| 193 |
-
l'audio sur toute sa durée. Il est exprimé en **dBFS** (decibels relative to
|
| 194 |
-
full scale : 0 dBFS = pleine échelle), donc toujours ≤ 0 ; plus la valeur est
|
| 195 |
-
**proche de 0**, plus l'audio est **fort/chargé en énergie**.
|
| 196 |
-
|
| 197 |
-
Calcul, sur le WAV rendu complet `x` (échantillons dans `[-1, 1]`) :
|
| 198 |
|
| 199 |
-
|
| 200 |
-
|
| 201 |
```
|
| 202 |
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
**
|
| 208 |
-
|
| 209 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 210 |
|---|---|
|
| 211 |
-
|
|
| 212 |
-
|
|
| 213 |
-
|
|
|
|
|
|
|
|
| 214 |
|
| 215 |
-
|
| 216 |
-
|---|---|---|---|
|
| 217 |
-
| STEAK-rare | 419,716 | 420,218 | 839,934 |
|
| 218 |
-
| STEAK-medium | 420,434 | 419,479 | 839,913 |
|
| 219 |
-
| STEAK-tough | 419,864 | 419,983 | 839,847 |
|
| 220 |
|
| 221 |
-
|
| 222 |
-
paire est **toujours entièrement dans `train` OU entièrement dans `test`**, jamais
|
| 223 |
-
un rôle dans `train` et l'autre dans `test` — aucune fuite atco↔pilot entre les
|
| 224 |
-
deux splits. Cela vaut **quel que soit le subset** : les deux rôles peuvent tomber
|
| 225 |
-
dans des subsets de difficulté différents (`STEAK-rare`/`-medium`/`-tough`, selon
|
| 226 |
-
leur bruit propre), mais ils partagent **toujours le même split**.
|
| 227 |
|
| 228 |
-
|
|
|
|
| 229 |
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
|
| 233 |
-
|
| 234 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 235 |
|
| 236 |
-
|
| 237 |
|
| 238 |
-
|
| 239 |
-
vocalisation type *uh / um / er*). Dans la transcription `text`, cette hésitation
|
| 240 |
-
est marquée par le token **`[hes]`**, et la colonne **`hesitation`** en donne le
|
| 241 |
-
**type** (`null` si l'énoncé n'en contient pas) :
|
| 242 |
|
| 243 |
-
|
|
|
|
|
|
|
|
|
|
| 244 |
|---|---|
|
| 245 |
-
|
|
| 246 |
-
|
|
| 247 |
-
|
|
| 248 |
-
|
|
| 249 |
-
|
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
| 253 |
-
|
| 254 |
-
-
|
| 255 |
-
|
| 256 |
-
|
| 257 |
-
|
| 258 |
-
|
| 259 |
-
|
| 260 |
-
|
| 261 |
-
|
| 262 |
-
|
| 263 |
-
|
| 264 |
-
|
| 265 |
-
-
|
| 266 |
-
|
| 267 |
-
|
| 268 |
-
|
| 269 |
-
|
| 270 |
-
|
| 271 |
-
|
| 272 |
-
`atco` et `pilot` d'un échange partagent `pair_id`. Comme leur niveau de bruit
|
| 273 |
-
diffère, les deux rôles peuvent tomber dans des subsets différents : pour
|
| 274 |
-
reconstituer les paires complètes, charge les trois subsets.
|
| 275 |
|
| 276 |
-
|
| 277 |
-
from collections import defaultdict
|
| 278 |
-
from datasets import load_dataset, concatenate_datasets
|
| 279 |
|
| 280 |
-
|
| 281 |
-
|
| 282 |
-
|
| 283 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 284 |
|
| 285 |
-
|
| 286 |
-
for ex in ds:
|
| 287 |
-
pairs[ex["pair_id"]][ex["role"]] = ex # {'atco': ..., 'pilot': ...}
|
| 288 |
|
| 289 |
-
|
| 290 |
-
|
| 291 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
path: data/STEAK-tough/test-*.parquet
|
| 33 |
---
|
| 34 |
|
| 35 |
+
# STEAK — **S**peech-to-**T**ext for **E**rror of **A**tc readbac**K**
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
|
| 37 |
+
STEAK is a **synthetically generated** dataset of **ATCO–pilot radio exchanges**
|
| 38 |
+
— both the **text** and the **audio** are synthetic:
|
| 39 |
|
| 40 |
+
- **Text**: assembled by **formal rules**, following an **ontology of ATCO–pilot
|
| 41 |
+
exchanges**.
|
| 42 |
+
- **Audio**: **TTS** → voice **timbre / accent** conversion (**seed-vc**) →
|
| 43 |
+
**noise** addition (noise captured from real **ATCO2** recordings).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
|
| 45 |
+
**One row = one audio** (one ATCO controller utterance *or* one pilot readback).
|
| 46 |
+
**2,519,694 audios.** The **ATCO↔pilot pair is always present**: both roles of an
|
| 47 |
+
exchange share the same `pair_id`. To rebuild the complete pairs directly:
|
| 48 |
|
| 49 |
+
```python
|
| 50 |
+
from collections import defaultdict
|
| 51 |
+
from datasets import load_dataset, concatenate_datasets
|
| 52 |
|
| 53 |
+
subsets = ["STEAK-rare", "STEAK-medium", "STEAK-tough"]
|
| 54 |
+
ds = concatenate_datasets(
|
| 55 |
+
[load_dataset("DEEL-AI/STEAK", s, split="train") for s in subsets]
|
| 56 |
+
)
|
| 57 |
|
| 58 |
+
pairs = defaultdict(dict)
|
| 59 |
+
for ex in ds:
|
| 60 |
+
pairs[ex["pair_id"]][ex["role"]] = ex # {'atco': ..., 'pilot': ...}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 61 |
|
| 62 |
+
pair = pairs["0000001"]
|
| 63 |
+
atco, pilot = pair.get("atco"), pair.get("pilot")
|
| 64 |
```
|
| 65 |
|
| 66 |
+
## Text generation (formal rules + ATC ontology)
|
| 67 |
+
|
| 68 |
+
82 commands in total can be generated (full table below). An utterance
|
| 69 |
+
contains **at most 7** (1 *seed* command + up to 6 added) and **2.2 on
|
| 70 |
+
average** (1 seed + ~1.2 added). The **first (seed) command** is drawn at random
|
| 71 |
+
among the 82 with a **weighting** (not all equally likely): the weights
|
| 72 |
+
come from **EDA observations** on the real ATC corpora **ATCO2** and
|
| 73 |
+
**UWB-ATCC**, so that the en-route core (climb/descend/heading/contact…)
|
| 74 |
+
dominates and rare commands stay rare. **Multiple phrasings per command**: each
|
| 75 |
+
command has several controller-side phrasings, and several ways to be **read
|
| 76 |
+
back** on the pilot side — each phrasing is drawn at random with a weighting
|
| 77 |
+
obtained from our EDA observations on ATCO2 and UWB-ATCC. The readback phrasing
|
| 78 |
+
is **not necessarily** the same as the instruction (sometimes reduced to the
|
| 79 |
+
value alone, *“level 100”*).
|
| 80 |
+
|
| 81 |
+
We also added **incompatibility rules** within a single utterance: at most **one
|
| 82 |
+
command per group** (never two vertical commands, nor two heading commands,
|
| 83 |
+
etc.), vertical-direction coherence (no *“climb … rate of descent”*),
|
| 84 |
+
ground/flight exclusivity (no *“taxi … descend FL90”*) and phase conflicts
|
| 85 |
+
(departure ⊥ arrival…).
|
| 86 |
+
|
| 87 |
+
The concrete values (waypoints, station names & frequencies, airports,
|
| 88 |
+
airlines/callsigns) come from **public sources**:
|
| 89 |
+
|
| 90 |
+
| resource | source |
|
| 91 |
|---|---|
|
| 92 |
+
| waypoints | [OpenNav](https://opennav.com/waypoint/FR) (one page per country) |
|
| 93 |
+
| ICAO airports | [OpenNav](https://opennav.com/airportcodes) |
|
| 94 |
+
| ATC stations + frequencies | [VATEUD](https://fsmine.dhis.org/vateud8/) |
|
| 95 |
+
| airlines / callsigns (telephony) | [Wikipedia](https://en.wikipedia.org/wiki/List_of_airline_codes), [flugzeuginfo.net](https://www.flugzeuginfo.net/), [OpenNav](https://opennav.com/airlinecodes/), [FAA CNT](https://www.faa.gov/air_traffic/publications/atpubs/cnt_html/) |
|
| 96 |
+
| seed-vc reference voices (accent, `accent_ref`) | [Speech Accent Archive](https://accent.gmu.edu) |
|
| 97 |
|
| 98 |
+
### The 82 commands and their value domains
|
|
|
|
|
|
|
|
|
|
|
|
|
| 99 |
|
| 100 |
+
The 82 commands are grouped **by family** (vertical, heading, speed…).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 101 |
|
| 102 |
+
<details>
|
| 103 |
+
<summary>Show the 82 commands and their value domains</summary>
|
| 104 |
|
| 105 |
+
| command type | value domain |
|
| 106 |
+
|---|---|
|
| 107 |
+
| Vertical — climb / descend / level (climb, descend, maintain, stop climb/descent, rate of climb/descent…) | flight level **FL060–FL400** (steps of 10) or altitude **2000–15000 ft** (steps of 1000) ; vertical rate **500–3000 ft/min** |
|
| 108 |
+
| Approach & landing (cleared to land, cleared ILS / VOR / NDB / DME / visual approach, touch-and-go) | **runway 01–36** (+ suffix **L / R / C**) |
|
| 109 |
+
| Waypoint & route navigation (direct to, inbound, cross waypoint ; follow route / STAR arrival / SID / transition) | a single **waypoint** (OpenNav), optionally + FL, for direct-to / inbound / cross ; **or** a whole **named published route** for follow-route — named after its reference fix: a **SID** (standard departure), a **STAR** (standard arrival), or a **transition** (the segment linking an en-route point to a SID/STAR) |
|
| 110 |
+
| Ground runway movements (cross runway, hold short, line up, backtrack, vacate) | **runway 01–36** (+ suffix **L / R / C**) ; for *vacate*, a **taxiway id** (a phonetic letter, optionally + a number or a 2nd letter — e.g. `A`, `L`, `A48`) |
|
| 111 |
+
| Heading (turn left/right, turn by N°, fly heading) | an **absolute heading 005–360°** (multiples of 5 — *“fly heading 230”*) ; **or** a **relative turn** left/right of **{10, 20, 30, 40, 90, 180, 270, 360}°** (*“turn right by 30 degrees”*) |
|
| 112 |
+
| QNH | **975–1045 hPa** |
|
| 113 |
+
| Transponder (squawk, ident) | **0000–7777 octal** code (emergency/reserved codes 7500/7600/7700/7777/0000 excluded) ; *ident* takes no value |
|
| 114 |
+
| Speed (speed, reduce / increase speed, maintain speed, mach) | **160–340 kt** (steps of 10), or **Mach 0.70–0.86** (steps of 0.01) |
|
| 115 |
+
| Frequency transfer (contact, monitor) | **station + frequency 118.000–136.xxx MHz** (source VATEUD) |
|
| 116 |
+
| Ground & departure (startup, pushback, taxi to / via, hold position) | a **taxiway id** (phonetic letter, optionally + number — `A`, `A48`) ; no value for *startup / pushback / hold position* |
|
| 117 |
+
| Time constraint (time) | **hhmm time** (+ waypoint) |
|
| 118 |
+
| Holding | a **waypoint** (OpenNav) |
|
| 119 |
+
| Information (ATIS, active runway, traffic, wind, destination, CTR entry) | ATIS **alpha…zulu** ; active runway **01–36** ; traffic as a relative position (« x o'clock y miles »…) ; wind **direction / speed / gust** ; destination **airport (ICAO, OpenNav)** ; **CTR entry waypoint** |
|
| 120 |
+
| Replies & initiation (affirm, standby, go ahead, initial call) | — (no value) |
|
| 121 |
|
| 122 |
+
</details>
|
| 123 |
|
| 124 |
+
A command may carry a **condition** — a temporal/spatial trigger attached to it.
|
|
|
|
|
|
|
|
|
|
| 125 |
|
| 126 |
+
<details>
|
| 127 |
+
<summary>The 26 condition connectors (« … » = generated value)</summary>
|
| 128 |
+
|
| 129 |
+
| connector | attaches to commands |
|
| 130 |
|---|---|
|
| 131 |
+
| « when ready » | direct-to / inbound, frequency contact, vertical (climb / descend / maintain…) |
|
| 132 |
+
| « when able » | direct-to / inbound, vertical (climb / descend / maintain…) |
|
| 133 |
+
| « when established » | approach, vertical (climb / descend / maintain…) |
|
| 134 |
+
| « when passing … feet » | vertical (climb / descend / maintain…) |
|
| 135 |
+
| « until further advised » | heading, speed, vertical (climb / descend / maintain…) |
|
| 136 |
+
| « until passing flight level … » | vertical (climb / descend / maintain…) |
|
| 137 |
+
| « until … » | direct-to / inbound, vertical (climb / descend / maintain…) |
|
| 138 |
+
| « until … miles » | speed, vertical (climb / descend / maintain…) |
|
| 139 |
+
| « after … » | direct-to / inbound, heading, vertical (climb / descend / maintain…) |
|
| 140 |
+
| « after … then … » | direct-to / inbound, heading, vertical (climb / descend / maintain…) |
|
| 141 |
+
| « after the low approach » | heading, vertical (climb / descend / maintain…) |
|
| 142 |
+
| « after departure » | direct-to / inbound, vertical (climb / descend / maintain…) |
|
| 143 |
+
| « level by … » | vertical (climb / descend / maintain…) |
|
| 144 |
+
| « before … » | direct-to / inbound, vertical (climb / descend / maintain…) |
|
| 145 |
+
| « at … feet per minute » | vertical (climb / descend / maintain…) |
|
| 146 |
+
| « at … feet per minute or greater » | vertical (climb / descend / maintain…) |
|
| 147 |
+
| « time … » | cross waypoint, direct-to / inbound, vertical (climb / descend / maintain…) |
|
| 148 |
+
| « from present » | direct-to / inbound, heading, vertical (climb / descend / maintain…) |
|
| 149 |
+
| « from present position » | direct-to / inbound, heading |
|
| 150 |
+
| « if able » | direct-to / inbound, speed, vertical (climb / descend / maintain…) |
|
| 151 |
+
| « if possible » | direct-to / inbound, speed, vertical (climb / descend / maintain…) |
|
| 152 |
+
| « if available » | vertical (climb / descend / maintain…) |
|
| 153 |
+
| « until … miles final » | approach, speed, vertical (climb / descend / maintain…) |
|
| 154 |
+
| « on conversion » | speed |
|
| 155 |
+
| « when established on the localizer » | approach, speed, vertical (climb / descend / maintain…) |
|
| 156 |
+
| « when you are done de icing » | frequency contact |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 157 |
|
| 158 |
+
</details>
|
|
|
|
|
|
|
| 159 |
|
| 160 |
+
**At most 2 conditions per utterance.** The probability of having a condition in
|
| 161 |
+
an utterance is **0.08**, and of having 2 conditions
|
| 162 |
+
**0.08 × 0.18 = 0.0144**. The command
|
| 163 |
+
the condition is applied to is drawn at random, and on the pilot side the
|
| 164 |
+
condition is read back with probability **0.55**. The
|
| 165 |
+
condition appears in the `commandes` column (`atco_condition`) and, if read back,
|
| 166 |
+
in `pilot_condition_readback`.
|
| 167 |
+
|
| 168 |
+
## Audio rendering: voice, accent, noise
|
| 169 |
+
|
| 170 |
+
The text is then synthesized and degraded to resemble a real ATC frequency:
|
| 171 |
+
|
| 172 |
+
- **Voice / timbre**: each utterance is synthesized by **Kokoro TTS**, with a
|
| 173 |
+
**voice drawn at random** among the Kokoro voices (column `voice`).
|
| 174 |
+
- **Accent (voice conversion)**: seed-vc then transfers the voice of a reference
|
| 175 |
+
clip onto the synthesis. The reference clips (column `accent_ref`) are scraped
|
| 176 |
+
from the [Speech Accent Archive](https://accent.gmu.edu) (≈ 3000 speakers of
|
| 177 |
+
all native languages reading the same paragraph → varied accents). The pool is
|
| 178 |
+
split by gender and the share of male voices is drawn per role (≈ 0.7 ATCO /
|
| 179 |
+
≈ 0.9 pilot).
|
| 180 |
+
- **Noise**: a noise profile from **real ATCO2 recordings** is added (column
|
| 181 |
+
`atco2_noise_profile`). No reverberation (small, absorptive ATC cabins →
|
| 182 |
+
negligible T60).
|
| 183 |
+
|
| 184 |
+
<video controls width="100%" src="https://huggingface.co/datasets/DEEL-AI/STEAK/resolve/main/media/noise_chain.mp4">
|
| 185 |
+
Your browser does not support the video tag — <a href="https://huggingface.co/datasets/DEEL-AI/STEAK/resolve/main/media/noise_chain.mp4">download it</a>.
|
| 186 |
+
</video>
|
| 187 |
+
|
| 188 |
+
## Difficulty (measured-RMS tercile)
|
| 189 |
+
|
| 190 |
+
Difficulty = tercile of the **RMS measured on the rendered WAV** (column `rms`).
|
| 191 |
+
|
| 192 |
+
The **RMS** (*root-mean-square*) is the **effective signal level**: the average
|
| 193 |
+
energy/volume of the audio over its whole duration. It is expressed in **dBFS**
|
| 194 |
+
(decibels relative to full scale: 0 dBFS = full scale), so always ≤ 0; the
|
| 195 |
+
**closer to 0**, the **louder / more energetic** the audio. Since the audio is
|
| 196 |
+
**peak-normalized** at the channel output (peak fixed at 0 dBFS), the RMS — which
|
| 197 |
+
includes the background noise and the soft-clip fold-back — is **comparable
|
| 198 |
+
across files**. **High RMS = noisy audio = harder.**
|
| 199 |
+
|
| 200 |
+
| subset | rms bounds (dBFS) |
|
| 201 |
+
|---|---|
|
| 202 |
+
| `STEAK-rare` (easy, clean) | `<= -13.08` |
|
| 203 |
+
| `STEAK-medium` | `]-13.08, -11.55]` |
|
| 204 |
+
| `STEAK-tough` (hard, noisy) | `> -11.55` |
|
| 205 |
|
| 206 |
+
## Train / Test split
|
|
|
|
|
|
|
| 207 |
|
| 208 |
+
| subset | train | test | total |
|
| 209 |
+
|---|---|---|---|
|
| 210 |
+
| STEAK-rare | 419,716 | 420,218 | 839,934 |
|
| 211 |
+
| STEAK-medium | 420,434 | 419,479 | 839,913 |
|
| 212 |
+
| STEAK-tough | 419,864 | 419,983 | 839,847 |
|
| 213 |
+
|
| 214 |
+
The `train`/`test` split (~50% test) is cut **by pair** (`pair_id`): a
|
| 215 |
+
pair is **always entirely in `train` OR entirely in `test`**, never one role in
|
| 216 |
+
`train` and the other in `test` — no atco↔pilot leakage between the two splits.
|
| 217 |
+
This holds **regardless of subset**: the two roles may fall into different
|
| 218 |
+
difficulty subsets (`STEAK-rare`/`-medium`/`-tough`, depending on their own
|
| 219 |
+
noise), but they **always share the same split**.
|
media/noise_chain.mp4
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:199190d46b13129b45bcd1f565ae09ae325f1f056e8e94adccc5b0c39bcfdafb
|
| 3 |
+
size 11130973
|