twangodev/radiotalk-us-transcripts-grok-4.20-50k

radiotalk-us-transcripts-grok-4.20-50k

49,984 synthetic US air-traffic-control transcripts, generated with xAI's grok-4.20-0309-non-reasoning against the v2 radiotalk scenario pipeline. Built for fine-tuning ATC ASR models (NVIDIA Parakeet, Whisper, etc.) and for seeding TTS audio generation.

Third release in the radiotalk transcripts series, and the first from a non-Qwen generator:

• —v1: `twangodev/radiotalk-us-transcripts-qwen3-100k`
• —v2: `twangodev/radiotalk-us-transcripts-qwen3-25k`
• —v3: this dataset

What's different vs v2

Same scenario machinery, different generator. The v2 release's FAA-aligned class distributions, OpenSky-ADS-B-weighted operator sampling, 86-hub airport restriction, real runway lists, ARTCC injection, service-airframe compatibility rules, prompt p2, and the post-generation realism validator all carry over unchanged — see the v2 card for the full write-up. What changed:

• —Generator: grok-4.20-0309-non-reasoning via the xAI API (OpenAI-compatible chat completions), no reasoning/thinking mode, temperature 0.9, in place of local Qwen/Qwen3-32B-NVFP4 with thinking enabled.
• —Scale: 49,984 transcripts vs 22,065 — a fresh seed-42 scenario stream, so scenarios do not overlap v2's.
• —Validator pass behavior: the same reject-and-retry validator that drove an 11.7% reject rate on Qwen3-32B rejected well under 1% of Grok 4.20 first attempts; only 68 of ~51,500 attempted scenarios (0.13%) exhausted retries and were dropped.

Distribution snapshot

operator_class:  commercial 41.9 / ga 30.1 / business 10.1 / cargo 5.8 /
                 training 4.9 / military 4.0 / rotorcraft 2.0 / medevac 1.1
phase:           tower 29.9 / approach 25.1 / ground 19.9 / center 15.0 / ramp 10.0
event:           routine 83.5 / abnormal 12.5 / emergency 4.0
weather:         VMC 81.7 / IMC 18.3
time of day:     day 65.4 / night 24.7 / dawn 5.0 / dusk 4.9
density:         moderate 49.6 / heavy 30.5 / light 19.9
airports:        86 unique US tier-1+2 hubs, well-balanced (~2.4% each at top)
size:            527,820 turns, ~45.7M utterance characters

Schema

Identical to v2 (including the artcc field) — see the v2 card for the full column and struct reference. The model column reads grok-4.20-0309-non-reasoning on every row.

Generation

• —Model: grok-4.20-0309-non-reasoning (xAI API), no reasoning mode
• —Concurrency: 32 in-flight requests, ~12 transcripts/sec sustained
• —Wall time: ~72 minutes across two legs (seed 42, resumable manifest)
• —Prompt p2, taxonomy t1, validator as shipped in v2

Companion datasets

• —`twangodev/radiotalk-voices-2k` — 2,000 reference voices for TTS cloning
• —`twangodev/radiotalk-us-audio-higgs-clean` / `-noisy` — v2 audio via Higgs Audio v2
• —`twangodev/radiotalk-us-audio-tada-clean` / `-noisy` — v1 audio via Hume TADA

Quality evals

• —LLM-judge realism scores (`judge_scores.jsonl`): every transcript scored 1-10 by grok-4.20-0309-non-reasoning against an FAA-phraseology rubric at temperature 0. Mean 6.7; distribution 3: 6 / 4: 542 / 6: 24,691 / 7: 13,152 / 8: 11,506; 1.1% score below 5. The judge is deliberately strict and flags some intentional design choices (e.g. ICAO "decimal" vs FAA "point"), so treat scores as a relative quality signal and outlier filter.
• —Closed-loop STT check: Grok STT on clean TTS renderings of these transcripts reaches 5.8% mean WER / 75% exact-match — see `radiotalk-us-audio-grok-clean`.

License

CC-BY-4.0. Attribute the radiotalk project + xAI for the underlying LLM. Operator-frequency weights derived from the OpenSky ADS-B archive under their academic-use terms.