---
license: mit
language:
- en
tags:
- speech-recognition
- asr
- ctc
- conformer
- p25
- imbe
- vocoder
- fine-tuned
- onnx
datasets:
- librispeech_asr
- LIUM/tedlium
- speechcolab/gigaspeech
base_model: trunk-reporter/imbe-asr-base-512d
pipeline_tag: automatic-speech-recognition
---

# IMBE-ASR Base P25 Fine-tuned (48.6M params)

P25 radio-adapted variant of [imbe-asr-base-512d](https://huggingface.co/trunk-reporter/imbe-asr-base-512d). Produces readable transcriptions from real P25 radio traffic.

**Code:** [trunk-reporter/imbe-asr](https://github.com/trunk-reporter/imbe-asr) | **Base model:** [imbe-asr-base-512d](https://huggingface.co/trunk-reporter/imbe-asr-base-512d) | **Best model:** [imbe-asr-large-1024d](https://huggingface.co/trunk-reporter/imbe-asr-large-1024d)

## Results

Evaluated on 50 real P25 labeled samples using greedy decode vs. beam search with the included 3-gram KenLM:

| Decode method | WER | CER |
|---|---|---|
| Greedy | 37.1% | 14.8% |
| Beam + KenLM (α=0.5, β=1.0) | **19.2%** | **9.5%** |

The KenLM reduces WER by ~18 percentage points. **Beam search with the included LM is strongly recommended.**

Example P25 output: `BATTALION 60 ENGINE 62 MEDIC 61 RESPOND TO 1234 MAIN STREET FOR A MEDICAL EMERGENCY`

## Training

Fine-tuned from `imbe-asr-base-512d` on ~20 hours of real P25 radio captures, pseudo-labeled with Whisper large-v3 + Qwen3-ASR ensemble. Mixed with 30% base training data to prevent catastrophic forgetting.

## Files

| File | Format | Size |
|------|--------|------|
| `model.safetensors` | SafeTensors | 205 MB |
| `config.json` | JSON | — |
| `model.onnx` | ONNX fp32 | 196 MB |
| `model_int8.onnx` | ONNX int8 | 58 MB |
| `stats.npz` | NumPy | 2 KB |
| `lm/3gram.bin` | KenLM trie (3-gram, q8) | 501 MB |
| `lm/unigrams.txt` | Vocabulary | 9 MB |

## Usage

### Greedy decode (fast, no dependencies)

```python
import onnxruntime as ort, numpy as np

session = ort.InferenceSession("model_int8.onnx")
stats = np.load("stats.npz")
features = ((raw_params - stats["mean"]) / stats["std"]).astype(np.float32)
log_probs, out_lengths = session.run(None, {
    "features": features.reshape(1, -1, 170),
    "lengths": np.array([features.shape[0]], dtype=np.int64),
})
```

### Beam search + KenLM (recommended, ~18pp WER improvement)

```python
import onnxruntime as ort, numpy as np
from pyctcdecode import build_ctcdecoder
import kenlm

# Load model
session = ort.InferenceSession("model_int8.onnx")
stats = np.load("stats.npz")

# Build decoder with KenLM — tuned params for P25
VOCAB = list(" ABCDEFGHIJKLMNOPQRSTUVWXYZ'")  # 28 chars + blank at index 0
labels = [""] + VOCAB
decoder = build_ctcdecoder(
    labels=labels,
    kenlm_model_path="lm/3gram.bin",
    unigrams=open("lm/unigrams.txt").read().splitlines(),
    alpha=0.5,   # LM weight — tuned on P25 data
    beta=1.0,    # word insertion bonus
)

# Run inference
features = ((raw_params - stats["mean"]) / stats["std"]).astype(np.float32)
log_probs, out_lengths = session.run(None, {
    "features": features.reshape(1, -1, 170),
    "lengths": np.array([features.shape[0]], dtype=np.int64),
})
text = decoder.decode(log_probs[0, :out_lengths[0]], beam_width=100)
```

Install dependencies: `pip install pyctcdecode kenlm`

## Limitations

- Pseudo-labeled training data may contain transcription errors.
- P25 coverage is primarily law enforcement, fire, and EMS from one region. May not generalize to all agencies.
- A P25 fine-tuned version of the large-1024d model is in progress and will substantially outperform this one.
- English only.