Voice Enhancement in Noisy Vehicle Intercom Using ADAU1463

Hello everyone,

I’d like to ask the community for support with a technical challenge involving audio processing in a high-noise environment. I'm developing a vehicle intercom system whose purpose is to allow clear communication between people outside the vehicle and those inside the vehicle, particularly in urban environments with high ambient noise (traffic, engines, wind, etc.).

System Objective
The main goal is for someone inside the vehicle to clearly hear what a person outside is saying, even when there is significant environmental noise. The system is intended for use cases such as access control, roadside assistance, or interaction with external personnel without lowering windows.

Hardware and Setup

Main DSP: ADAU1463

Currently using 2 microphones, expandable up to 4

Sampling rate: 16 kHz

Entire audio processing chain developed in SigmaStudio

Microphone Placement and Topologies Tested
We’ve explored different microphone configurations to capture speech in this noisy scenario:

Spatially separated microphones:
Two microphones were placed with significant distance between them, aiming for one to capture primarily environmental noise (reference input) and the other to capture voice + noise. This setup was used with adaptive filtering and subtraction algorithms.

Beamforming topology:
We implemented beamforming using the available SigmaStudio block. Two directional beams were formed (left and right sides of the vehicle), mimicking “digital ears.” VAD blocks were used to detect which beam had stronger speech content, and the main mic was selected dynamically based on energy and VAD flags.

We're still assessing which topology works better. If anyone in the community has experience with this kind of setup, recommendations on microphone placement and topology are welcome.

Important: In all cases, the speaker is roughly 1 meter away from the microphones, so this is not close-range voice pickup, and environmental noise strongly affects all channels.

Nature of the Noise
The environment is highly noisy, and this noise affects all microphones equally. It includes:

Vehicle engines

Urban traffic (horns, footsteps, wind, structural vibrations)

Unpredictable and non-stationary noise sources

This makes separating speech from background noise very challenging without specialized processing.

Algorithms and Techniques Tried So Far

High-pass and band-pass filtering

VAD and VAD flags — though loud noises often trigger false positives

NLMS adaptive filter, using a noise reference and desired input (voice + noise), with the error as the output

Multiband analysis (6 bands), using energy comparison and dynamic gating logic

Spectral subtraction algorithms, following Speech Enhancement, 2nd Edition:

Classical spectral subtraction

Spectral subtraction with over-subtraction and spectral flooring

Nonlinear spectral subtraction

Attempted implementation of multiband spectral subtraction, but the per-bin alpha estimation was too complex to manage in SigmaStudio’s frequency-domain blocks. A simplified time-domain version was tested and worked well in quiet environments, but still left considerable residual noise in real-world conditions.

Trial NS block (link) from SigmaStudio was also tested, but did not yield useful results under our scenario.

What We're Looking For
We would greatly appreciate help from the community on the following:

Recommendations on microphone topology: How many microphones are realistically needed, and where should they be placed to best capture voice in this context?

Viable algorithms: Are there any noise reduction or voice enhancement strategies that can be realistically implemented on the ADAU1463 using SigmaStudio?

Best practices for speech capture at ~1 meter distance in high-noise outdoor environments

Beamforming and source separation insights that could work effectively within SigmaStudio’s limitations

Most importantly, we’re looking for guidance on which path is actually worth pursuing — to avoid wasting time and resources exploring microphone configurations or algorithm types that are unlikely to work in practice.

A clear recommendation of an effective topology and processing strategy would be extremely helpful.

The scenario may sound complicated (and it is), but if anyone in the community has worked on similar applications or can share insights, we’d be deeply grateful.

Thank you in advance for your time and support!

Audio Signal Processors

hardware

adau1463

SigmaDSP Audio Processors