Skip to content

Telink Edge AI Noise Suppression using TL721x User Guide

Overview

Limitations of traditional voice noise suppression methods

Methods and limitations: the traditional voice noise suppression (also known as noise reduction) methods mainly include statistical modeling-based methods and signal processing-based methods, such as Wiener filtering and spectral subtraction. These methods are usually based on some simple assumptions and fixed algorithms, which have better results to a certain extent for specific types of noise, however their noise suppression performance is often limited in complex and variable noise environments. They are difficult to accurately model the complex characteristics of voice and noise, especially for non-smooth noise and noise with time-varying characteristics, and are prone to excessive or incomplete noise suppression, and may also introduce side effects such as musical noise, which affects the naturalness and intelligibility of voice.

Advantages of Deep Learning Technology

Deep Learning Advantage: the deep learning has powerful nonlinear modeling ability, feature learning ability and data fitting ability. It can automatically learn the complex features and patterns of voice and noise from a large amount of data without designing complex feature extractors and noise suppression algorithms manually. By constructing a deep neural network model, it can represent and process various information in voice signals more comprehensively and accurately, thus showing great potential in voice noise suppression tasks.

Hardware and Software

Hardware: TL721X platform motherboard AIOT-DK1, TL7218D based module ML7218D-MERCURY-M0-PE11-V01, CODEC daughterboard AIOT-CODEC1, stereo 3.5mm wired headphone, USB Type-C cable

Software: nn_ns_demo.bin (this bin is for neural network noise suppression demonstration)

The hardware connection is as below:

Hardware connection of the demonstration

Download Program

According to the figure below, install the ML7218D-MERCURY-M0-PE11-V01 module and the CODEC daughterboard AIOT-CODEC1, and using Dupont cables to connect the motherboard AIOT-DK1 to the programmer. Burn the above bin file in the motherboard AIOT-DK1.

Download program to motherboard AIOT-DK1

Operation Description

The motherboard AIOT-DK1 is configured with keys and two LEDs for controlling and displaying the algorithm in different states. The SW2 is for resetting the development board, SW4 for selecting the noise suppression channel mode, and SW6 for selecting the algorithm’s noise suppression depth. The location of the keys and LEDs are shown below

Keys and LEDs on the board

  • LED D2: by default the yellow-green light is always on at power-on, indicating that the program is running.

  • LED D1: by default the white light is off at power-on, indicating that AI noise suppression is not enabled.

  • SW2: the switch of resetting the development board.

  • SW4: the switch of selecting the noise suppression channel mode.

    • Dual-Channel Noise Suppression: In this mode, the yellow-green light is always on at power-on.

    • Single-Channel Noise Suppression: In this mode, the yellow-green light is always off at power-on.

  • SW6: the switch of controlling AI noise suppression, and controlling the cycle switch between the following noise suppression depths, the cycle order is 0dB -> 20dB -> 25dB -> 30dB -> unlimited -> 0dB:

    • Threshold 1: 0dB (algorithm does not take effect), the white light is always off, it is in this state at power-on by default;

    • Threshold 2: 20dB, the white light blinks once in 1s;

    • Threshold 3: 25dB, the white light blinks once in 600ms;

    • Threshold 4: 30dB, the white light blinks once in 300ms;

    • Threshold 5: unlimited, the white light is always on.

Noise suppression effect

When AI noise suppression is not turned on, using the headphones we can hear obvious noise from the surrounding environment. When AI noise suppression is turned on, ambient noise is clearly eliminated.