RoamingPigs/trcs-tactical-audio

Open Tactical Radio Channel Simulator (TRCS) Benchmark

The Open Tactical Radio Channel Simulator (TRCS) benchmark is an open evaluation harness and acoustic dataset designed to characterize speech model hallucination, narrow-band codec degradation, and tactical entity extraction under realistic defense, maritime, and public safety communications channels.

Drawing directly from operational US Navy, Coast Guard (USCG), and Department of Homeland Security (DHS) communications profiles, TRCS provides standardized evaluation across:

1. Narrowband Telephony Codecs: G.711 $\mu$-law, G.711 A-law, 3GPP AMR-NB, ETSI GSM 06.10 Full-Rate, and IETF Opus Narrowband.
2. RF Channel Transceiver Physics: Squelch tail RF bursts, squelch opening/closing transients, CTCSS sub-audible tone injection, and calibrated SNR additive mixing.
3. Operational Evaluation Splits:
   • hallucination (E1): Non-speech silence, ambient room noise, and RF squelch bursts measuring insertion rates (WPM and clip hallucination fraction).
   • telephone_grid (E2): Tactical Search-and-Rescue (SAR) cutter and helicopter radio transmissions scored on Word Error Rate (WER) and Tactical Entity Error Rate (EER).

---

1. Formal Acoustic Specifications

Channel Degradation	Specification / Standard	Implementation Parameters	Operational Physical Meaning
Bandwidth Limiting	ITU-T G.712 telephone speech bandpass	300 Hz – 3400 Hz (4th-order Butterworth)	Legacy copper PSTN & military tactical audio circuits
G.711 $\mu$-law (PCMU)	ITU-T G.711 ($\mu=255$)	8 kHz sampling, 8-bit logarithmic PCM (64 kbps)	North American PSTN, marine VHF DSC gateways
G.711 A-law (PCMA)	ITU-T G.711 ($A=87.6$)	8 kHz sampling, 8-bit logarithmic PCM (64 kbps)	International maritime telephony & European PSTN
AMR-NB	3GPP TS 26.071 (MR122)	8 kHz sampling, ACELP vocoder (12.2 kbps)	Tactical cellular, satellite telephony, push-to-talk
GSM 06.10 FR	ETSI GSM 06.10 Full-Rate	8 kHz sampling, RPE-LTP vocoder (13.0 kbps)	Combat net cellular, legacy tactical GSM
Opus Narrowband	IETF RFC 6716	8 kHz sampling, SILK mode (12.0 kbps VoIP)	Tactical mesh radios, MANET tactical IP networks
Squelch Tail Burst	FM Discriminator Unmuting	50–150 ms duration, bandpass filtered hiss	RF carrier drop before receiver audio gate closes
CTCSS Sub-audible Tone	EIA/TIA-603 standard	100.0 Hz / 131.8 Hz / 151.4 Hz (-18 dBc)	Continuous Tone-Coded Squelch repeater access tone
SNR Noise Mixing	Calibrated $P_{\text{signal}} / P_{\text{noise}}$	$+20\text{ dB}, +10\text{ dB}, 0\text{ dB}, -6\text{ dB}$	Pink, brown, white, turbine engine, babble noise

---

2. Tactical Channel Profiles

The benchmark includes 7 canonical pre-configured channels (TACTICAL_PROFILES):

1. clean_control: 16 kHz uncompressed studio baseline (no codec, no noise).
2. pstn_g711: Standard 8 kHz G.711 $\mu$-law at 30 dB SNR white thermal noise.
3. tactical_vhf_clear: G.711 $\mu$-law at 20 dB SNR pink noise, 80 ms squelch tail, 100.0 Hz CTCSS tone.
4. tactical_vhf_fringe: AMR-NB at 5 dB SNR pink noise, 120 ms squelch tail, 131.8 Hz CTCSS tone.
5. tactical_gsm: GSM Full-Rate at 15 dB SNR pink noise, 60 ms squelch tail.
6. mesh_radio_opus: Opus Narrowband at 15 dB SNR multi-speaker ambient babble.
7. extreme_combat_engine: AMR-NB at -2 dB SNR diesel/gas-turbine engine noise, 150 ms squelch tail, 151.4 Hz CTCSS.

---

3. Evaluation Splits & Structure

Split 1: hallucination (E1 Benchmark)

Evaluates speech recognition models on non-speech acoustic audio to measure false insertion rate.

• Conditions: digital_silence, ambient_silence (-55 dBFS), squelch_burst, noise_white, noise_pink, noise_brown, noise_engine, noise_babble, and speech_control.
• Target Metrics:
  • Clip Hallucination Rate: $\frac{N_{\text{hallucinated}}}{N_{\text{non-speech clips}}}$.
  • Words Per Minute (WPM): Total inserted words normalized to audio duration.

Split 2: telephone_grid (E2 Benchmark)

Evaluates speech recognition on operational search-and-rescue (SAR) cutter and helicopter radio transmissions through the full 6-codec $\times$ 4-SNR acoustic matrix.

• Target Metrics:
  • Word Error Rate (WER): Standard Levenshtein dynamic programming alignment.
  • Tactical Entity Error Rate (EER): Precision, Recall, and F1 score on tactical callsigns (e.g. Cutter Bertholf CG-7501, Cutter Munro CG-755, Rescue 6033), tactical VHF channels (e.g. Channel 16, 22 Alpha), and NATO phonetic coordinates (Sierra Bravo, Yankee Zulu 89).

---

4. Quickstart: 2-Line Evaluation Harness

You can evaluate any Hugging Face checkpoint or custom speech function with two lines of code:

from data.trcs_hf import TRCSEvaluator

# Score any open speech model checkpoint on tactical WER and entity error rate:
results = TRCSEvaluator.evaluate("openai/whisper-tiny.en", split="telephone_grid")
TRCSEvaluator.print_summary(results)

To evaluate a custom transcription callable or pipeline:

# Custom callable: func(audio_path) -> str
results = TRCSEvaluator.evaluate(my_asr_function, split="telephone_grid")

# Evaluate hallucination split:
hallucination_results = TRCSEvaluator.evaluate(my_asr_function, split="hallucination")
print(f"Hallucination Rate: {hallucination_results['hallucination_clip_rate']:.1%}")
print(f"Inserted Words/Min: {hallucination_results['words_per_minute_hallucinated']:.1f}")

---

5. Loading via Hugging Face datasets

from datasets import load_dataset

# Load complete tactical telephone grid
ds_grid = load_dataset("data/trcs_hf", "telephone_grid", split="test")
print(ds_grid[0])
# {
#   'id': 'e2_cg_01_g711_mulaw_snr20',
#   'text': 'Coast Guard Sector Columbia River, this is Cutter Bertholf CG-7501...',
#   'codec': 'g711_mulaw',
#   'snr_db': 20.0,
#   'entities': ['bertholf', 'cg-7501', '16', 'sierra', 'bravo'],
#   'audio_path': 'audio/e2_cg_01_g711_mulaw_snr20.wav'
# }

# Load hallucination benchmark split
ds_hal = load_dataset("data/trcs_hf", "hallucination", split="test")

---

6. Dataset Schema

Each example contains:

• id (string): Unique clip identifier.
• audio_path (string): Path to the 16 kHz WAV audio file.
• text (string): Reference ground-truth transcript (empty for non-speech audio).
• entities (list[string]): Scripted tactical entities (callsigns, hulls, channels, phonetics).
• condition (string): Channel profile or noise condition.
• codec (string): Codec applied (none, g711_mulaw, g711_alaw, amr_nb, gsm_fr, opus_nb).
• snr_db (float32): Signal-to-noise ratio in decibels (null for uncorrupted clean control).
• duration_sec (float32): Audio duration in seconds.
• is_speech (bool): Boolean flag indicating whether speech is present.

---

7. Citation

@dataset{caceres2026trcs,
  title   = {Tactical Radio Channel Simulator (TRCS): Robustness & Hallucination Benchmark},
  author  = {Caceres, Cisco},
  year    = {2026},
  url     = {https://ciscocaceres.com/research}
}