---
license: apache-2.0
pipeline_tag: audio-to-audio
tags:
- fullsubnet
- speech-enhancement
- audio
- robotics
- IB-Robot
- ascend
- torch
- edge-deployment
---

# Model Card for FullSubNet (IB-Robot)

FullSubNet speech enhancement model with cumulative Laplace normalization for
streaming 4-channel microphone array processing, packaged for the
[IB-Robot](https://atomgit.com/openeuler/IB_Robot) framework with three
deployments sharing one stream contract.

## Deployments

| deployment | backend | artifacts | notes |
|---|---|---|---|
| `ascend_310p` | Ascend ACL (Ascend310P1) | stateful FB/SB OM pair | board-side streaming (LSTM state via state_links) |
| `torch_cuda` | PyTorch CUDA | — (ckpt in `assets/`) | host-side streaming executor |
| `torch_cpu` | PyTorch CPU | — (ckpt in `assets/`) | CPU fallback (watch the 128 ms hop budget) |

All deployments share `tensor_model/fullsubnet/enhance`
(`observation.audio_4ch [-1,4] -> voice.audio_enhanced_4ch [-1,4]`) with a
stateful stream execution contract (`state_bank_mode: runtime_exclusive`,
`max_open_streams: 1`).

## Weights provenance

- `assets/cum_fullsubnet_best_model_218epochs.tar` = Audio-WestlakeU/FullSubNet
  official v0.2 release checkpoint (sha256
  `d08d09107eb276b8dc3d2d9fff995f4354a51fa3347125f52f8b9aea7c339f81`)
- The 310P stateful OM pair was converted from the same checkpoint
- `assets/cum_fullsubnet_best_model_218epochs.manifest.json` pins the digest and
  `norm_type: cumulative_laplace_norm` (do not mix with offline-norm checkpoints)

Note: the `.tar` file is the upstream PyTorch serialization container
(legacy `torch.save` format), loaded directly by `torch.load` — no extraction
step.

## Repository Structure

- `inference_manifest.json` — deployment routing (schema v3, stateful stream contract)
- `assets/adapter.json` — algorithm contract (STFT 512/256, T=2, look-ahead 2)
- `assets/cum_fullsubnet_best_model_218epochs.tar` — Torch checkpoint
- `artifacts/ascend/fullsubnet/*.om` — stateful FB/SB OM modules

## Usage

Host: `voice_asr_service.speech_direction` with backend `stateful_torch_cuda` /
`stateful_torch_cpu`; board: select `ascend_310p` via the unified inference
runtime.