twangodev/radiotalk-us-audio-tada-noisy

RadioTalk US Audio (Noisy)

VHF AM aviation channel-degraded variants of synthesized US air-traffic-control speech. One row per (clean turn × variant), embedded 8 kHz mono PCM_16 WAV.

This is the noisy variant of twangodev/radiotalk-us-audio-tada-clean — same transcripts and voices, passed through a probabilistic channel-simulation pipeline calibrated to the ATCO2 corpus SNR distribution (mean ~8 dB, range -5 to +30 dB) and shaped to ITU-R M.1084 / DO-186B aero voice passband (300-3400 Hz, 6th-order Butterworth).

Quick start

from datasets import load_dataset
ds = load_dataset("twangodev/radiotalk-us-audio-tada-noisy", split="train", streaming=True)
row = next(iter(ds))
print(row["text"], row["voice_id"], row["profile"], row["effective_snr_db"])
print(row["audio"]["array"].shape, row["audio"]["sampling_rate"])

Stats

• 3,365,754 noisy rows = 1,121,918 clean utterances × 3 independent channel realizations
• 8 kHz mono PCM_16 WAV (matches ATCO2-PL, the canonical aviation-ASR rate; bandpass output has no content above 4 kHz)
• 674 parquet shards (~470 MB each), audio bytes embedded inline (audio.bytes)

Schema

Strict superset of tada-clean. Anyone reading the clean dataset can read this one with no code changes; the extra fields are additive.

Passthrough from clean

Field	Type	Description
scenario_id	str	sha1 of scenario definition
scenario	struct	icao, region, phase, aircraft[], weather, ... (full clean struct)
turn_idx	int64	turn position within the scenario
speaker	str	speaker label (ATC facility or aircraft callsign)
text	str	raw transcript
text_normalized	str	ATC-normalized transcript (numbers spelled out, etc.)
voice_id	str	TADA reference voice id
audio	struct	8 kHz mono PCM_16 WAV bytes (degraded)
tokens	list[TokenSpan] | null	per-token timing from TADA (clean-side timing; not re-aligned to noisy audio)
model	str	upstream transcript-generation LLM
prompt_version	str	upstream prompt version
taxonomy_version	str	upstream scenario taxonomy version
tts_model	str	upstream TTS model (HumeAI/tada-3b-ml)

Radio-pipeline additions

Field	Type	Description
clean_row_id	str	{scenario_id}#{turn_idx} — links this variant back to its clean source
variant_idx	int64	0, 1, or 2 — which of the 3 variants this is
seed	int64	deterministic per-(clean_row_id, variant_idx) RNG seed
profile	str	pilot or controller — sampled per variant, drives effect probabilities
applied_effects	list[str]	exact effect chain applied to this row, e.g. ["bandpass(300-3400,6)", "pink_noise(snr=8.2)", "key_click", "squelch_tail"]
effective_snr_db	float64	measured signal-to-noise ratio after additive noise
pipeline_version	str	semantic version of the channel-sim code that produced this row
pipeline_fingerprint	str	16-char hash of the full profile config (effects + probs + SNR sampler + sample_rate) for that variant

Channel-simulation pipeline

For each variant, the clean WAV is passed through a probabilistic chain. Effects are gated independently with per-profile probabilities and the order is fixed:

clean 24 kHz FLAC
    -> decode + resample to 8 kHz
    -> [agc_pump]               (p=0.10-0.15)
    -> [tanh saturation]        (p=0.30-0.40, drive 3-15 dB)
    -> [hard clip / hot mic]    (p=0.08-0.15, percentile 95-99)
    -> bandpass 300-3400 Hz     (always)
    -> [multipath flutter]      (pilot only, p=0.25)
    -> pink noise at SNR        (always, target SNR sampled from
                                 ATCO2-calibrated Gaussian: mean=8,
                                 std=9, clipped [-5, +30] dB)
    -> bandpass 300-3400 Hz     (re-applied per Su & Haq 2026 recipe)
    -> [heterodyne whistle]     (p=0.03; "stepped on" beat tone)
    -> [PTT key click]          (p=0.70-0.75)
    -> [squelch tail]           (p=0.75-0.80)
    -> [codec roundtrip]        (p=0.30, currently Ogg Vorbis)
    -> peak-limit to 0.99

Pilot-side draws SNR from a noisier-uplink sampler (mean 4 dB) and has multipath flutter enabled; controller-side draws from a clean-tower sampler (mean 14 dB) and skips flutter. Profile is sampled 50/50 per variant.

Reproducibility

• Variant seed = sha256(f"{scenario_id}#{turn_idx}|{variant_idx}")[:8] masked to 63 bits.
• Profile, SNR, all effect gates and parameters are drawn from a numpy.random.default_rng(seed).
• Bytes are not guaranteed identical across pipeline_version changes; always join via clean_row_id + variant_idx, not row position.

Sources

• Clean audio: twangodev/radiotalk-us-audio-tada-clean
• Transcripts: twangodev/radiotalk-us-transcripts-100k
• Voices: twangodev/radiotalk-voices-2k
• Pipeline code: https://github.com/twangodev/radiotalk

License

Built with Llama.

The audio in this dataset was synthesized using HumeAI/tada-3b-ml, which is built on Meta Llama 3.2 3B, then passed through a deterministic DSP channel simulator. As such, this dataset is distributed under the Llama 3.2 Community License Agreement.

Llama 3.2 is licensed under the Llama 3.2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.

See the full license text at https://www.llama.com/llama3_2/license/.