---
title: speech_jaec_aec_16k
canonical_url: "https://www.modelscope.cn/models/iic/speech_jaec_aec_16k"
md_url: "https://www.modelscope.cn/models/iic/speech_jaec_aec_16k.md"
repository: iic/speech_jaec_aec_16k
chinese_name: "端到端回声消除-jaec-16k-base"
last_updated: 2026-09-09
license: "Apache License 2.0"
pipeline_tag: acoustic-echo-cancellation
tasks:
  - acoustic-echo-cancellation
model_type:
  - jaec
library_name:
  - pytorch
frameworks:
  - PyTorch
domain:
  - audio
downloads: 26
stars: 30
tags:
  - Alibaba
  - JAEC
  - "acoustic echo cancellation"
  - AEC
  - "speech enhancement"
  - "real-time communication"
---

# speech_jaec_aec_16k

> speech_jaec_aec_16k - iic 在 ModelScope 开源的模型。JAEC 是一个用于实时通信的 16 kHz 声学回声消除前端。作为可解释型端到端AEC的此版本包含神经网络时延估计（TDE）和自适应线性处理（LP）运行时。它接收麦克风信号和远端参考信号，并返回 LP 输出。

iic/speech_jaec_aec_16k 是 ModelScope 魔搭社区上的acoustic-echo-cancellation模型，采用 Apache License 2.0 许可。

- **Repository**: iic/speech_jaec_aec_16k
- **License**: Apache License 2.0
- **Tasks**: acoustic-echo-cancellation
- **Tags**: Alibaba, JAEC, acoustic echo cancellation, AEC, speech enhancement, real-time communication
- **Downloads**: 26
- **Stars**: 30
- **Last updated**: 2026-09-09

Source: https://www.modelscope.cn/models/iic/speech_jaec_aec_16k

---

# JAEC

JAEC is a 16 kHz acoustic echo cancellation frontend for real-time
communication. This release contains the neural time-delay estimation (TDE)
and adaptive linear processing (LP) runtime. It accepts microphone and far-end
reference signals and returns the LP output.

The native runtime processes 160 samples (10 ms) per call. The output retains
a fixed algorithmic delay of 352 samples (22 ms) relative to the microphone
input.

![JAEC TDE and LP frontend architecture](assets/jaec_architecture.png)

## Usage

JAEC requires **ModelScope > 1.39.1**. Once a compatible release is available on PyPI, install it with:

```bash
python -m pip install -U "modelscope>1.39.1"
```

Until then, install ModelScope from the official `master` branch:

```bash
python -m pip install -U \
  "modelscope @ git+https://github.com/modelscope/modelscope.git@master"
```

Run inference with two 16 kHz, mono, PCM16 WAV files of equal length:

```python
from modelscope.outputs import OutputKeys
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks


aec = pipeline(
    Tasks.acoustic_echo_cancellation,
    model='iic/speech_jaec_aec_16k',
    device='cpu',
    trust_native_code=True,
)

result = aec(
    {
        'nearend_mic': 'https://dashscope.oss-cn-beijing.aliyuncs.com/samples/audio/jaec/nearend_mic.wav',
        'farend_speech': 'https://dashscope.oss-cn-beijing.aliyuncs.com/samples/audio/jaec/farend_speech.wav',
    },
    output_path='output.wav',
)
pcm16_bytes = result[OutputKeys.OUTPUT_PCM]
```

`nearend_mic` and `farend_speech` may also be WAV bytes or one-dimensional
`numpy.int16` arrays. The pipeline pads an incomplete final 160-sample frame
internally and trims the output back to the original length.

The model repository contains native libraries. Set `trust_native_code=True`
only when loading a repository that you trust.

## Supported platforms

| Operating system | Architecture | Native library | Runtime requirement |
| --- | --- | --- | --- |
| macOS | Arm64 | `lib/jaec_arm.so` | macOS 11 or later |
| Linux | x86-64 | `lib/jaec_x86.so` | AVX2, FMA, glibc 2.27 or later |
| Windows | x86-64 | `lib/jaec_x86.dll` | 64-bit Windows |

The included Arm64 library is a macOS Mach-O binary. It does not support Linux
Arm64, Android, or iOS. The package does not include a 32-bit runtime.

## Runtime performance

Each RTF is the median of 10 runs. The timed region covers Python per-frame
inference and excludes WAV I/O and model initialization.

| Operating system | Architecture | CPU | Main clock | RTF |
| --- | --- | --- | --- | ---: |
| macOS | Arm64 | Apple M4 | 4.46 GHz | 0.0020 |
| Linux | x86-64 | Intel Xeon | 2.90 GHz | 0.0039 |
| Windows | x86-64 | Intel Core Ultra 9 185H | 2.30 GHz | 0.0029 |

## Limitations

- Inputs must be equal-length 16 kHz mono PCM16 signals, with the microphone in
  `nearend_mic` and the far-end reference in `farend_speech`.
- The output is the JAEC TDE+LP frontend result; this package does not include
  the nonlinear processing (NLP) network.
- Runtime compatibility is limited to the operating systems and architectures
  listed above.

## License

This release is licensed under the
[Apache License 2.0](LICENSE). Third-party attributions required by the native
binaries are provided in
[`THIRD_PARTY_NOTICES.md`](THIRD_PARTY_NOTICES.md).