Short answer. The easy half of this problem — suppressing acoustic noise at the microphone, before encoding — is occupied, by Motorola and Hytera in the radio itself and by DVSI inside the vocoder, and they are doing it competently. The hard half — repairing speech that has already been through a 2,450 bps parametric vocoder and a bad RF link — is close to unoccupied, is technically feasible on a sub-watt part, and is where a generative model will confidently invent the wrong street name. The project is not blocked by compute or by data. It is blocked by hallucination under negative SNR, by the fact that the background noise is itself operational information, and by a certification and procurement moat that has very little to do with machine learning.
1 · Who is already here
Break the signal path into four places a cleanup stage can live, because the incumbents are not spread evenly across them.
⟨A⟩ Pre-vocoder, in the radio — crowded and competent
Motorola's APX audio engine is the reference point. The XVP remote speaker microphone carries four digital microphones and runs adaptive noise suppression, adaptive gain, adaptive speaker equalization, and adaptive windporting, which brings a third microphone in specifically to cancel wind. Motorola's own claim is that the talker stays audible over diesel engines, power equipment and sirens; the acoustic design — port geometry, array placement so the talker can speak from any direction — is doing as much work as the DSP. Hytera's H-series, announced March 2025, markets AI-based noise suppression on the same premise.
DVSI does a version of this inside the codec. AMBE+2 v1.4 (adopted 2005) improved tone handling; v1.6 (adopted 2009) added noise rejection aimed at high-frequency background, explicitly to stop a firefighter's PASS alarm from destroying the voice model.PASS = Personal Alert Safety System, the ~95 dB motion-triggered alarm on a firefighter's SCBA. It is a loud, tonal, high-frequency signal sitting right on top of speech — close to a worst case for a harmonic vocoder. NIST reported improvement in noisy conditions from that version. This matters for the proposal: the most valuable single intervention has already been made, once, by the vocoder vendor, and it shipped seventeen years ago.
⟨A⟩ The hardware answer — the strongest competitor, and it isn't ML
INVISIO, Savox, 3M Peltor and Silynx solve the noise problem by not letting the noise into the microphone. The INVISIO X5 puts the microphone in the ear and picks the voice off jawbone vibration; the same earpiece gives 32 dB SNR / 29 dB NRR of hearing protection. Bone conduction has no background-noise problem to speak of, works at a whisper, and is in service with the Canadian Army's Integrated Soldier System. Any ML pitch has to beat "put the transducer somewhere the noise isn't," which is a physical solution with no failure mode, no latency, and no hallucination.
⟨C⟩ At the console — newly occupied, thinly
Motorola Solutions now ships AI noise reduction in Guardian ICCS, its dispatch and telephony software for UK and European emergency services: a toggle in the call handler's UI, quoted at 16 ms of added latency, published April 2026. That is a fair marker of where the incumbent thinks the accessible value is — the receiving end, in software, where there is no radio to recertify.
⟨B⟩ Post-vocoder, in the earpiece or speaker — effectively open
The nearest things to prior art are adjacent rather than competing. Polyn Technology announced a NeuroVoice chip (December 2025) and a NeuroComm headset for two-way radios using neuromorphic analog signal processing — voice activity detection under 30 µW, voice extraction to 8 kHz, sub-millisecond inference, headset total under 200 µW — which is a remarkable power envelope but is aimed at activation and extraction, not at repairing vocoded speech. On the amateur side, RMnoise sits between a receiver and a sound card and denoises SSB and CW audio; it is a hobbyist tool with no published metrics. Argonne and collaborators assembled ~62,000 manually transcribed Chicago Police Department transmissions (~46 hours) and benchmarked Whisper large-v3 and custom ESPnet models — that is ASR on radio audio, which shares the domain but solves a different problem.
Verdict on occupancy. Nobody is selling a model that takes decoded, degraded P25 audio and reconstructs intelligible speech from it in real time. That gap is real. It is also partly a gap because the incumbents have concluded — reasonably — that once the vocoder has discarded the parameters, there is nothing left to recover honestly.
2 · Why it sounds like that
Two separate failure modes get blamed on the same thing. Analog radios sound bad because of noise. Digital radios sound bad because of the model. They need different answers.
The channel is nothing like a cell channel
| LMR (P25) | Cellular (LTE/NR) | |
|---|---|---|
| Channel width | 12.5 kHz | 5–100 MHz |
| Voice payload | 2,450–4,400 bps | 13,200–24,000 bps typical |
| Audio band | ~300–3,400 Hz | 50–7,000 Hz (AMR-WB), to 20 kHz (EVS) |
| Site spacing | tens of km | 0.5–3 km urban |
| Power control | none in direct mode | closed loop, ~1,500×/s |
| Retransmission | none — one shot | HARQ, ARQ |
| Latency budget | ~200–500 ms, half duplex | <150 ms, full duplex |
| Must work with no network | yes (talkaround) | no |
The last row is the one that explains most of the others. A phone that loses the network stops being a phone. A radio that loses the network has to keep working simplex, portable to portable, through a building, which forces a design that is robust rather than good: low bit rate so the link closes at the edge of coverage, no retransmission because there is no return path in a one-to-many push-to-talk call, and a narrow channel because the spectrum allocation is 12.5 kHz and has been since the FCC narrowbanding mandate.
Why analog hisses
FM demodulation has a noise power spectrum that rises with frequency — the "noise triangle" — so the hiss you hear is weighted toward exactly the band where fricatives live. Pre-emphasis at the transmitter and de-emphasis at the receiver exist to fight this, at the cost of headroom. FM also has the capture effect: the stronger of two signals takes the channel entirely, which is why a distant unit can be erased by a closer one rather than merely competing with it. Multipath on a moving vehicle produces flutter and picket-fencing at a rate set by speed and wavelength, and CTCSS squelch tails add the characteristic chirp at the end of a transmission. All of this is additive, stationary-ish, and — importantly — a denoiser can help, because the waveform is still there underneath.
Why digital warbles
P25 does not transmit speech. It transmits the parameters of a speech model and resynthesizes at the far end. Phase 1 uses IMBE at 4,400 bps of voice plus 2,800 bps of FEC — 7,200 bps on air in a 9,600 bps C4FM channel. Phase 2 halves that: two TDMA slots in the same 12.5 kHz, AMBE+2 half-rate at 2,450 bps voice plus 1,150 bps FEC. Every 20 ms the encoder estimates a pitch, a voiced/unvoiced decision per frequency band, and a spectral envelope, and ships those numbers.
Three consequences follow directly, and together they are the entire "squawky" sound:
- Model mismatch. The model assumes one harmonic source. A siren is harmonic. A PASS alarm is harmonic. Two people talking are two harmonic sources. When the input violates the assumption, the encoder fits the interferer, and the decoder faithfully resynthesizes something that was never speech. Noise does not get added to the voice — it replaces the voice, in the voice's own timbre.
- The cliff. FEC corrects errors up to a threshold and then stops. Past it, the decoder repeats the last good frame or mutes. Analog degrades continuously and a listener can lean in; digital is intelligible, intelligible, gone.
- Tandeming. Audio crossing a console, a patch, an analog gateway or an interoperability bridge is decoded and re-encoded. Each pass re-parameterizes material that is already synthetic. The second encoder is now modelling the first decoder's artifacts.
There is also a physiological term nobody controls. Under high noise a talker raises level and shifts spectrum — the Lombard effect — and shouted, strained speech is precisely the regime where harmonic estimation is least reliable. The conditions that create the noise also create the input the vocoder is worst at.
Why this is still true in 2026
Not inertia alone, though there is plenty of that in a market with 15–25 year radio lifecycles and sole-source procurement. The binding constraints:
- The spectrum didn't get wider
- 12.5 kHz channels are a regulatory allocation. Every bit of quality has to come out of coding efficiency, not bandwidth.
- Interoperability is the product
- P25 exists so that a mutual-aid unit from the next county can key up on your system. The vocoder is frozen into the standard and P25 CAP testing; changing it means every radio in the region, at once.
- The vocoder is licensed IP
- IMBE and AMBE+2 are DVSI's, patent-encumbered and delivered as tested binaries. This is a large part of why open experimentation in this area has stayed at the hobbyist edge.
- The device constraints are brutal
- Intrinsic safety approval for explosive atmospheres, MIL-STD-810, IP68, a 12-hour shift on one battery, and operation in a helmet and gloves. Every milliwatt is contested.
The military side shows the same curve at a lower bit rate: MELPe under MIL-STD-3005 and STANAG-4591 at 2,400 / 1,200 / 600 bps, with a 300 bps MELP device built for DARPA in 2010 by MIT Lincoln Laboratory, Compandent, BBN and General Dynamics — better than 600 bps MELPe but at longer algorithmic delay. The trade has always been bits against delay against robustness, and ML changes the shape of that trade without removing it.
3 · Feasibility of a real-time neural cleanup layer
Assume arbitrary data and training compute, as posed, and custom silicon on the device. The question then splits by where the layer sits, because the constraint sets are not comparable.
| Point | Whose hardware | Latency you can spend | Real obstacle |
|---|---|---|---|
| ⟨A⟩ pre-vocoder | Motorola / Hytera / Tait | ~20 ms | you are not in the box; and the incumbent is good here |
| ⟨B⟩ post-vocoder, live | earpiece, RSM, dispatch headset | ~40–80 ms | hallucination at negative SNR |
| ⟨C⟩ replay / logger / ASR | console PC, server | seconds | almost none — this is the wedge |
What genuinely favors the idea
The latency budget is unusually generous. Push-to-talk is half duplex. There is no conversational echo loop inside a transmission, and P25 already spends hundreds of milliseconds on trunking grant, vocoder framing, interleaving and network transit. Adding 40 ms on the receive path is beneath the noise floor of what an operator perceives. DeepFilterNet's 40 ms of algorithmic latency — 20 ms window, 50% overlap, two-frame lookahead — would be a hard constraint in a telephony product and is free here.
The compute is not the problem. Low-complexity speech enhancement at these rates already runs continuously in hearing aids and earbuds on sub-watt DSPs (Cadence HiFi, Arm Ethos-U, Syntiant, Ambiq). Polyn's numbers — hundreds of microwatts, microsecond inference — show the floor is far below what a radio battery cares about. "Custom hardware" is not required; it is an optimization.
You can synthesize perfect paired data. This is the strongest argument and it is specific to this domain. Almost every speech-enhancement effort is starved of matched (clean, degraded) pairs. Here the degradation is a known, simulable pipeline: clean studio speech → measured fireground / siren / wind / SCBA impulse responses and noise → FM channel model or a P25 bitstream → injected bit errors at chosen BER → decoded audio. Open implementations of the MBE family exist for the analysis side, and the RF channel models are textbook. You can generate matched pairs at arbitrary scale, sweep SNR and BER as controlled variables, and never need a single field recording for training — only for validation.
Speaker personalization is available for free. Radios are issued to individuals. Ten seconds of enrolled voice turns blind denoising into target-speaker extraction, which is a much easier problem and simultaneously fixes cross-talk. No other deployment of speech enhancement gets a reliable speaker identity handed to it by the asset register.
The language is constrained. Unit identifiers, the phonetic alphabet, street grids, a bounded incident vocabulary. A domain language model over the output is worth a lot of SNR.
What could make it impossible
- The information is gone
- Below roughly 0 dB SNR, and after a parametric vocoder has thrown away everything but pitch, band voicing and envelope, the distinctions that separate five from nine or Tango from Charlie are not attenuated — they are absent. Enhancement is then generation. A model that outputs clean, confident speech from that input is not recovering anything; it is sampling a plausible continuation.
- Hallucination is the failure mode, and it is the unacceptable one
- The literature is unambiguous that diffusion and other generative enhancers hallucinate at negative input SNR — inpainting speech into noise-only regions, adding breaths and gasps, producing phonetic confusions. Recent work (predictive-then-generative cascades, phonological priors distilled from WavLM) reduces it and does not eliminate it. On a video call a hallucinated word is an annoyance. On a dispatch channel it is a wrong address, and the system has removed the very cue — the audible garbling — that would have made the listener ask for a repeat. An enhancement layer that hides its own uncertainty is worse than no layer.
- The noise is information
- A dispatcher hears the PASS alarm, the gunshots, the crowd, the breathing rate, the strain in the voice. Fireground practice explicitly treats background audio as situational awareness. Aggressive suppression deletes it; a generative resynthesis normalizes the talker's stress out of existence. This is not a tuning parameter — it is a product decision about what the channel is for, and getting it wrong will be rejected by users on the first shift.
- The output is evidence
- Radio traffic is logged, disclosed, and subpoenaed. Forensic audio practice requires processing that is reproducible, documented, and non-generative. A model that invents waveform cannot touch the archive. Any deployment has to treat the cleaned stream as a derived, clearly-labelled view with the original always one click away, and has to survive an admissibility challenge over what the model did.
- Evaluation is not where the ML field's metrics are
- PESQ, STOI and DNSMOS are not currency in this industry. The accepted instrument is intelligibility testing — the modified rhyme test, as used in NTIA/ITS TR-13-495 (Atkinson & Catellier, 2013) comparing analog FM and P25 in fireground noise, under the NIST public safety audio quality program. Generative models can improve perceptual-quality scores while lowering word accuracy. Budget for human MRT panels in simulated fireground noise, and test in loop — enhancement followed by re-encoding, since the vocoder may handle honest noise better than clean-sounding artifacts.
What makes it unlikely to work for a non-specialist
Almost none of this is machine learning.
- You cannot get into the box. ⟨A⟩ requires being the radio OEM or selling to one. The US LMR market is dominated by a single vendor with sole-source incumbency, multi-year contracts, and an accessory ecosystem it controls.
- Certification. Intrinsic safety (UL/FM/ATEX), MIL-STD-810, P25 CAP interoperability, and — if your processing sits inside the encryption boundary — a FIPS validation you do not want to pay for.
- Procurement. Agency purchasing runs on grant cycles, standards bodies and 3–7 year timelines. A startup's runway does not.
- Liability. If a generative layer renders an address incorrectly and someone is hurt, the vendor is the defendant. Indemnification for a probabilistic component in a life-safety path is not a solved commercial problem.
- Domain access. The validation data you need — real fireground audio, real SCBA speech, real PASS-alarm-masked transmissions — sits behind agencies, and the people who can evaluate whether your output is right rather than pleasant are firefighters and dispatchers, not annotators.
4 · What I would actually build
Enter at ⟨C⟩, where a human is in the loop and errors are correctable, and earn the right to move upstream.
- Instant replay and after-action. Every dispatch console already has a "replay last transmission" button. Make that replay clean. Non-real-time, non-safety- critical, human-verifiable, and it sells against a pain dispatchers describe unprompted.
- Two-stage, never one. A discriminative stage first — masking, deep filtering, no invention — then an optional generative refinement that is visibly marked and can be switched off. The cascade is what the hallucination literature recommends and it is also the only architecture that survives a deposition.
- Ship the uncertainty. Per-word confidence, and an audible or visual signal when the model is extrapolating. Design for the listener to know when to ask for a repeat — that is the actual safety function.
- Preserve the scene. Duck the background rather than deleting it; keep alarms and gunshots detectable; do not normalize vocal stress. Make this configurable per role — a dispatcher and an incident commander want different things.
- Measure with MRT, in loop. Human intelligibility panels against fireground noise at controlled SNR and BER, with and without downstream re-encoding. If MRT does not move, nothing else counts.
- Then ⟨B⟩. An earpiece or speaker-mic accessory, receive-side only, which needs no change to the radio, no change to the standard, and no OEM permission. That is the defensible product if the science holds.
The experiment that decides it. Build the simulation pipeline first — clean speech through measured fireground noise through a P25 encode/decode with injected bit errors. Train a discriminative enhancer on it. Then run an MRT panel at −5, 0, +5 and +10 dB SNR, with and without the model, both standalone and re-encoded. Two numbers decide the project: how much intelligibility the model adds where the signal still exists, and how often it produces a confident wrong word where it doesn't. The second number is the one that has killed this idea for everyone else.
Sources
- Motorola Solutions — XVP remote speaker microphone data sheet (4-mic array, adaptive noise suppression, windporting)
- Motorola Solutions — APX series audio quality fact sheet
- Motorola Solutions — AI audio enhancement in Guardian ICCS (April 2026, 16 ms)
- The AMBE+2 version 1.6 vocoder (v1.4 2005, v1.6 2009, PASS-alarm rejection, NIST testing)
- DVSI — benefits of end-to-end vocoder integration (tandeming)
- APCO Project 25 — RadioReference wiki (IMBE 4400+2800, Phase 2 half-rate 2450+1150)
- Atkinson & Catellier, NTIA/ITS TR-13-495 — Intelligibility of analog FM and updated P25 radio systems in fireground noise (modified rhyme test)
- NIST — Public Safety Audio Quality program
- INVISIO X5 bone conduction headset (32 dB SNR / 29 dB NRR)
- audioXpress — AI voice processing in two-way radios (Polyn NeuroVoice, NeuroComm, Dec 2025 / Feb 2026)
- Schröter et al. — DeepFilterNet and DeepFilterNet3 (40 ms algorithmic latency)
- Lemercier et al. — Diffusion models for audio restoration (hallucination at negative SNR)
- PASE — phonological priors for low-hallucination generative speech enhancement
- ALCF — speech recognition for police radio communication (~62,000 CPD transmissions, ~46 h)
- MELPe / MIL-STD-3005 / STANAG-4591 and the 2010 DARPA 300 bps MELP device
- RMnoise — AI noise reduction for radio audio