STEAK / README.md
camillebrl's picture
Mise à jour de la carte du dataset
571d897 verified
|
Raw
History Blame Contribute Delete
11.9 kB
metadata
license: mit
task_categories:
  - automatic-speech-recognition
language:
  - en
tags:
  - air-traffic-control
  - atc
  - synthetic-speech
size_categories:
  - 1M<n<10M
configs:
  - config_name: STEAK-rare
    default: true
    data_files:
      - split: train
        path: data/STEAK-rare/train-*.parquet
      - split: test
        path: data/STEAK-rare/test-*.parquet
  - config_name: STEAK-medium
    data_files:
      - split: train
        path: data/STEAK-medium/train-*.parquet
      - split: test
        path: data/STEAK-medium/test-*.parquet
  - config_name: STEAK-tough
    data_files:
      - split: train
        path: data/STEAK-tough/train-*.parquet
      - split: test
        path: data/STEAK-tough/test-*.parquet

STEAK — Speech-to-Text for Error of Atc readbacK

STEAK is a synthetically generated dataset of ATCO–pilot radio exchanges — both the text and the audio are synthetic:

  • Text: assembled by formal rules, following an ontology of ATCO–pilot exchanges.
  • Audio: TTS → voice timbre / accent conversion (seed-vc) → noise addition (noise captured from real ATCO2 recordings).

One row = one audio (one ATCO controller utterance or one pilot readback). 2,519,694 audios. The ATCO↔pilot pair is always present: both roles of an exchange share the same pair_id. To rebuild the complete pairs directly:

from collections import defaultdict
from datasets import load_dataset, concatenate_datasets

subsets = ["STEAK-rare", "STEAK-medium", "STEAK-tough"]
ds = concatenate_datasets(
    [load_dataset("DEEL-AI/STEAK", s, split="train") for s in subsets]
)

pairs = defaultdict(dict)
for ex in ds:
    pairs[ex["pair_id"]][ex["role"]] = ex   # {'atco': ..., 'pilot': ...}

pair = pairs["0000001"]
atco, pilot = pair.get("atco"), pair.get("pilot")

Text generation (formal rules + ATC ontology)

82 commands in total can be generated (full table below). An utterance contains at most 7 (1 seed command + up to 6 added) and 2.2 on average (1 seed + ~1.2 added). The first (seed) command is drawn at random among the 82 with a weighting (not all equally likely): the weights come from EDA observations on the real ATC corpora ATCO2 and UWB-ATCC, so that the en-route core (climb/descend/heading/contact…) dominates and rare commands stay rare. Multiple phrasings per command: each command has several controller-side phrasings, and several ways to be read back on the pilot side — each phrasing is drawn at random with a weighting obtained from our EDA observations on ATCO2 and UWB-ATCC. The readback phrasing is not necessarily the same as the instruction (sometimes reduced to the value alone, “level 100”).

We also added incompatibility rules within a single utterance: at most one command per group (never two vertical commands, nor two heading commands, etc.), vertical-direction coherence (no “climb … rate of descent”), ground/flight exclusivity (no “taxi … descend FL90”) and phase conflicts (departure ⊥ arrival…).

The concrete values (waypoints, station names & frequencies, airports, airlines/callsigns) come from public sources:

resource source
waypoints OpenNav (one page per country)
ICAO airports OpenNav
ATC stations + frequencies VATEUD
airlines / callsigns (telephony) Wikipedia, flugzeuginfo.net, OpenNav, FAA CNT
seed-vc reference voices (accent, accent_ref) Speech Accent Archive

The 82 commands and their value domains

The 82 commands are grouped by family (vertical, heading, speed…).

Show the 82 commands and their value domains
command type value domain
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
Approach & landing (cleared to land, cleared ILS / VOR / NDB / DME / visual approach, touch-and-go) runway 01–36 (+ suffix L / R / C)
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)
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)
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”)
QNH 975–1045 hPa
Transponder (squawk, ident) 0000–7777 octal code (emergency/reserved codes 7500/7600/7700/7777/0000 excluded) ; ident takes no value
Speed (speed, reduce / increase speed, maintain speed, mach) 160–340 kt (steps of 10), or Mach 0.70–0.86 (steps of 0.01)
Frequency transfer (contact, monitor) station + frequency 118.000–136.xxx MHz (source VATEUD)
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
Time constraint (time) hhmm time (+ waypoint)
Holding a waypoint (OpenNav)
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
Replies & initiation (affirm, standby, go ahead, initial call) — (no value)

A command may carry a condition — a temporal/spatial trigger attached to it.

The 26 condition connectors (« … » = generated value)
connector attaches to commands
« when ready » direct-to / inbound, frequency contact, vertical (climb / descend / maintain…)
« when able » direct-to / inbound, vertical (climb / descend / maintain…)
« when established » approach, vertical (climb / descend / maintain…)
« when passing … feet » vertical (climb / descend / maintain…)
« until further advised » heading, speed, vertical (climb / descend / maintain…)
« until passing flight level … » vertical (climb / descend / maintain…)
« until … » direct-to / inbound, vertical (climb / descend / maintain…)
« until … miles » speed, vertical (climb / descend / maintain…)
« after … » direct-to / inbound, heading, vertical (climb / descend / maintain…)
« after … then … » direct-to / inbound, heading, vertical (climb / descend / maintain…)
« after the low approach » heading, vertical (climb / descend / maintain…)
« after departure » direct-to / inbound, vertical (climb / descend / maintain…)
« level by … » vertical (climb / descend / maintain…)
« before … » direct-to / inbound, vertical (climb / descend / maintain…)
« at … feet per minute » vertical (climb / descend / maintain…)
« at … feet per minute or greater » vertical (climb / descend / maintain…)
« time … » cross waypoint, direct-to / inbound, vertical (climb / descend / maintain…)
« from present » direct-to / inbound, heading, vertical (climb / descend / maintain…)
« from present position » direct-to / inbound, heading
« if able » direct-to / inbound, speed, vertical (climb / descend / maintain…)
« if possible » direct-to / inbound, speed, vertical (climb / descend / maintain…)
« if available » vertical (climb / descend / maintain…)
« until … miles final » approach, speed, vertical (climb / descend / maintain…)
« on conversion » speed
« when established on the localizer » approach, speed, vertical (climb / descend / maintain…)
« when you are done de icing » frequency contact

At most 2 conditions per utterance. The probability of having a condition in an utterance is 0.08, and of having 2 conditions 0.08 × 0.18 = 0.0144. The command the condition is applied to is drawn at random, and on the pilot side the condition is read back with probability 0.55. The condition appears in the commandes column (atco_condition) and, if read back, in pilot_condition_readback.

Audio rendering: voice, accent, noise

The text is then synthesized and degraded to resemble a real ATC frequency:

  • Voice / timbre: each utterance is synthesized by Kokoro TTS, with a voice drawn at random among the Kokoro voices (column voice).
  • Accent (voice conversion): seed-vc then transfers the voice of a reference clip onto the synthesis. The reference clips (column accent_ref) are scraped from the Speech Accent Archive (≈ 3000 speakers of all native languages reading the same paragraph → varied accents). The pool is split by gender and the share of male voices is drawn per role (≈ 0.7 ATCO / ≈ 0.9 pilot).
  • Noise: a noise profile from real ATCO2 recordings is added (column atco2_noise_profile). No reverberation (small, absorptive ATC cabins → negligible T60).

Difficulty (measured-RMS tercile)

Difficulty = tercile of the RMS measured on the rendered WAV (column rms).

The RMS (root-mean-square) is the effective signal level: the average energy/volume of the audio over its whole duration. It is expressed in dBFS (decibels relative to full scale: 0 dBFS = full scale), so always ≤ 0; the closer to 0, the louder / more energetic the audio. Since the audio is peak-normalized at the channel output (peak fixed at 0 dBFS), the RMS — which includes the background noise and the soft-clip fold-back — is comparable across files. High RMS = noisy audio = harder.

subset rms bounds (dBFS)
STEAK-rare (easy, clean) <= -13.08
STEAK-medium ]-13.08, -11.55]
STEAK-tough (hard, noisy) > -11.55

Train / Test split

subset train test total
STEAK-rare 419,716 420,218 839,934
STEAK-medium 420,434 419,479 839,913
STEAK-tough 419,864 419,983 839,847

The train/test split (~50% test) is cut by pair (pair_id): a pair is always entirely in train OR entirely in test, never one role in train and the other in test — no atco↔pilot leakage between the two splits. This holds regardless of subset: the two roles may fall into different difficulty subsets (STEAK-rare/-medium/-tough, depending on their own noise), but they always share the same split.