Beijing Haiyitai Technology Co., Ltd.

1、Product Name & Model

Lightweight Real-time Speech Noise Reduction Algorithm

2、Product Image

Product Picture

3、Application Scenarios

Multi-terminal synchronous noise reduction processing; steady-state noise reduction for dynamic tracking engines, rotors and other noise sources; vehicle/airborne communication systems; tactical noise reduction headsets; helmet-mounted communication units; individual soldier combat communication terminals.

4、Main Purpose

Millisecond-level noise reduction, lightweight deployment, and clear voice transmission in full scenarios.

5、Core Performance

1)End-to-end processing latency: 20–30 ms;

2)Model parameters: ≤ 50K (via knowledge distillation and pruning);

3)Operating power consumption: 65–70 mA;

4)Supported sampling rates: mainstream sampling rates up to 48kHz;

5)Noise suppression capability: ≥35 dB for steady-state noise, ≥25 dB for non-steady-state noise;

6)Hardware adaptation: domestic military-grade chips;

7)Encryption algorithm support: domestic SM4 symmetric encryption algorithm (GB/T 32907-2016).

6、Core Technology & Advantages

Adopts deep learning time-frequency joint modeling to separate speech from steady-state and non-steady-state noise.A hybrid LSTM+CNN network architecture is applied to capture dual time-domain and frequency-domain features.The adaptive noise tracking algorithm enables dynamic adaptation to complex environments.Fixed-point quantization and SIMD parallel acceleration balance algorithm performance and power consumption.Dual optimization for military-grade requirements and localization: adaptable to a wide temperature range of -40℃~85℃, supports integrated domestic SM4 encrypted links, and is directly compatible with mainstream domestic chips without secondary development.