ASR directly from P25 IMBE codec parameters. Skip audio reconstruction, go straight from the digital bitstream to text. 1.9%% WER on LibriSpeech-IMBE.
AI & ML interests
None defined yet.
Recent Activity
View all activity
Organization Card
trunk-reporter
Open-source tools for P25 trunked radio transcription and monitoring.
We build speech recognition models and infrastructure for public safety radio -- law enforcement, fire, EMS, public works, transit, and more. Everything from real-time transcription to edge deployment on a Raspberry Pi.
Models
IMBE-ASR -- Speech Recognition from Vocoder Parameters
Skip audio reconstruction entirely. ASR directly from P25 IMBE codec parameters (4.4 kbps). 1.9% WER on LibriSpeech-IMBE with beam search.
| Model | Params | WER | Use Case |
|---|---|---|---|
| imbe-asr-large-1024d | 290M | 1.9% (beam) | Best accuracy, server/desktop |
| imbe-asr-base-512d | 48.6M | 18.9% (greedy) | Edge deployment (Pi 5, 15x RT) |
| imbe-asr-base-512d-p25 | 48.6M | -- | Real P25 dispatch fine-tuned |
All models available in SafeTensors, ONNX fp32, and ONNX int8/uint8 for edge deployment. The 290M model runs at 3.6x real-time on a Raspberry Pi 5 via a 70KB C binary.
Qwen3-ASR P25 -- Fine-tuned Audio ASR
| Model | Base | Use Case |
|---|---|---|
| qwen3-asr-p25-0.6B | Qwen3-ASR 0.6B | P25 audio transcription, pseudo-labeling |
Fine-tuned on P25 dispatch audio for conventional audio-based ASR. Used in our pseudo-labeling pipeline alongside Whisper large-v3.
Coming Soon
- IMBENet -- Neural IMBE vocoder for high-quality P25 audio reconstruction
Links
- GitHub
- trunk-recorder (fork with IMBE tap support)
models 4
trunk-reporter/imbe-asr-base-512d
Automatic Speech Recognition • Updated • 78
trunk-reporter/imbe-asr-large-1024d
Automatic Speech Recognition • Updated • 80
trunk-reporter/imbe-asr-base-512d-p25
Automatic Speech Recognition • Updated • 59
trunk-reporter/qwen3-asr-p25-0.6B
Automatic Speech Recognition • 0.8B • Updated • 35
datasets 0
None public yet