Lbrusset/SF_bay_voice2traj_dataset

SF Bay Voice2Traj Dataset

A multimodal dataset that aligns spoken Air Traffic Control (ATC) radio phrases with the corresponding aircraft trajectories. Each sample pairs a short audio clip of a controller/pilot transmission, its transcription, and the time-windowed ADS-B-style trajectory of the aircraft the phrase refers to.

The recordings cover ATC positions in the San Francisco Bay Area: San Francisco (KSFO), Oakland (KOAK), San Jose (KSJC) and Moffett Field (KNUQ), across delivery, ground, tower, approach and departure frequencies.

Key contribution. The raw audio and trajectory streams already exist as separate public sources. The value added by this dataset is the alignment that links each spoken radio phrase to the matching aircraft and its trajectory — matching transmissions to flights by callsign and time so that voice and motion become a single, jointly usable sample.

This dataset was built in an academic research context at the University of California, Berkeley (UC Berkeley), as part of the CITRIS research effort.

Overview

• Total samples: 52,030
• Audio source: LiveATC.net (licensed archives)
• Trajectory source: The OpenSky Network (open data)
• Audio: mono PCM, 16 kHz, int16
• Trajectories: fixed window of 14 points × 7 numeric features
• Format: HDF5 shards (one shard_XXXX.h5 per frequency)
• License: CC BY 4.0

Temporal Coverage

Date range: February 11–14, 2026 (about 3.5 days) Time window: 18:00 UTC (Feb 11) to 06:00 UTC (Feb 14)

• Evening / Night: 18:00–23:59 UTC
• Early morning: 00:00–06:00 UTC

This 60-hour window was selected to obtain high-quality, well-matched voice and trajectory samples across multiple ATC frequencies simultaneously.

Dataset structure

SF_bay_voice2traj_dataset/
├── KOAK-App/                      # Oakland Approach
│   └── shard_0000.h5
├── KOAK-Gnd-Twr/                  # Oakland Ground/Tower
│   └── shard_0000.h5
├── KSFO-KOAK-Dep/                 # SFO/KOAK Departure coordination
│   └── shard_0000.h5
├── KSFO-App-Combined/             # SFO Approach (28LR-only, BSR-MOD-Arr)
│   └── shard_0000.h5
├── KSJC-Dep/                      # San Jose Departure
│   └── shard_0000.h5
├── KSJC-App-Combined/             # San Jose Approach (App, App2)
│   └── shard_0000.h5
├── KNUQ-Gnd-Twr/                  # Moffett Ground/Tower
│   └── shard_0000.h5
└── README.md

Frequency	Samples	Notes
KOAK-App	3,706	Oakland Approach
KSFO-KOAK-Dep	21,102	SFO/KOAK Departure coordination
KSFO-App-Combined	12,953	SFO Approach (28LR-only, BSR-MOD-Arr)
KSJC-Dep	580	San Jose Departure
KSJC-App-Combined	5,875	San Jose Approach (App, App2)
KOAK-Gnd-Twr	7,701	Oakland Ground/Tower
KNUQ-Gnd-Twr	113	Moffett Ground/Tower
Total	52,030

Note: Some folder names use the -Combined suffix to indicate they contain samples from multiple sub-frequencies. These can be further subdivided if needed based on the metadata stored inside each HDF5 file.

HDF5 schema

Each shard_XXXX.h5 contains 4 aligned, index-parallel datasets of length N (number of samples in the shard):

Dataset	Shape	dtype	Description
waveforms	(N,)	vlen int16	Variable-length audio, 16 kHz mono PCM
metadata_json	(N,)	vlen str	Per-sample metadata as JSON
trajectories	(N, 14, 7)	float32	14 trajectory points × 7 numeric features
sample_indices	(N,)	int32	Stable sample id; -1 marks invalid/placeholder rows

Trajectory feature columns

0. time_position — UNIX timestamp (s)
1. longitude — degrees
2. latitude — degrees
3. geo_altitude — meters
4. velocity — m/s
5. true_track — degrees (0 = north, clockwise)
6. vertical_rate — m/s (may be NaN)

Per-sample metadata_json fields

{
  "sample_index": 0,
  "shard_idx": 0,
  "pos_in_shard": 0,
  "source_frequency": "koak_norcal_app",
  "icao24": "a52c4f",
  "callsign": "UAL667",
  "flight_start": "2026-02-11T23:59:54+00:00",
  "flight_end": "2026-02-12T00:19:55+00:00",
  "phrase_start": "2026-02-12T00:04:39.680+00:00",
  "phrase_end": "2026-02-12T00:04:41.200+00:00",
  "phrase_duration_s": 1.52,
  "transcription_text": "and then it's 667",
  "audio_sample_rate": 16000,
  "audio_n_samples": 36320,
  "trajectory": {
    "n_points": 14,
    "numeric_columns": ["time_position","longitude","latitude","geo_altitude","velocity","true_track","vertical_rate"],
    "icao24_per_point": ["a52c4f", "..."],
    "callsign_per_point": ["UAL667", "..."]
  }
}

Usage

import h5py, json, numpy as np

path = "KOAK-App/shard_0000.h5"

with h5py.File(path, "r") as f:
    n = f["trajectories"].shape[0]
    for i in range(n):
        if f["sample_indices"][i] == -1:
            continue  # Skip placeholder rows

        audio = np.asarray(f["waveforms"][i])  # int16, 16 kHz
        traj  = f["trajectories"][i]           # (14, 7) float32
        meta  = json.loads(f["metadata_json"][i])

        print(f"Sample {i}: {meta['callsign']} | {len(audio)} audio samples")

Data sources & attribution

• Audio (ATC voice): LiveATC.net — licensed ATC radio archives.
• Trajectories: The OpenSky Network — openly available aircraft surveillance (ADS-B) records.

The voice–trajectory alignment pipeline is documented in:

• CITRIS-joint-embedding-dataset-preprocessing
• CITRIS-llm-for-atc-dataset
• CITRIS-joint-dataset-analysis

Provenance & ethics

• Audio: Licensed from LiveATC.net. Original broadcasts are public ATC communications.
• Trajectories: Derived from The OpenSky Network, openly available aircraft surveillance records.
• Callsigns and ICAO 24-bit addresses are operational identifiers, not personal data.
• Users are responsible for complying with applicable laws regarding ATC/voice data use.

License

Released under Creative Commons Attribution 4.0 International (CC BY 4.0). You are free to share and adapt this dataset for any purpose, provided you give appropriate credit to the original sources (LiveATC.net, The OpenSky Network, and UC Berkeley).

Citation

@misc{sf_bay_voice2traj,
  title        = {SF Bay Voice2Traj Dataset},
  author       = {Louis Brusset},
  year         = {2026},
  howpublished = {UC Berkeley CITRIS},
  note         = {Audio licensed from LiveATC.net; trajectories from The OpenSky Network},
  url          = {https://github.com/LouisBrusset/CITRIS-joint-dataset-analysis}
}