Implementation of an Audio Reverberation Algorithm Fergal Dunne Supervisor: Dr. Edward Jones.

National University Ireland, GalwayofOllscoil na hÉireann, Gaillimh

Roinn na hInnealtóireachta LeichtreonaíDepartment Electronic Engineeringof

Implementation of an Audio Reverberation Algorithm

Fergal Dunne

Supervisor: Dr. Edward Jones



Project Aim

• To implement a reverberation algorithm for audio applications

• Gardner’s reverb algorithm - consists of three variations of the same general approach, intended to simulate rooms of different sizes

• Implementation of a real-time version• Simulation in Matlab -> C• Implementation on a DSP development

system/PC-Based system



Reverberation

• A multiplicity of echoes whose speed of repetition is too quick for them to be perceived as separate from one another

• Early and late reflections• Reverberation time – time

taken for the sound intensity to decay to 1/1,000,000th (60 dB) of its original value



Gardner’s Algorithm

• Combination of early and late echoes into a “unified” system, intended to simulate rooms of different sizes

• Early reflections trigger late reverberation network• Late reflections delayed until after the last early echo• Tap/gain included to produce the desired level of late

reverberation



Late Reflections

• Three similar algorithms – small, medium & large room• Common global structure based mainly, on allpass filters

(both simple and nested) and delays, followed by a first order low pass filter

• Output signal is obtained by summing signals taken from various points along the delay line



Which algorithm to use?

• The appropriate reverb algorithm is chosen based on the reverberation time of the room to be simulated (bathroom, auditorium…)

• Reverb time is calculated using Sabines Equation• Sabines equation uses information from the room to

be simulated - room dimensions, material, absorption coefficients…



Room Responses - Bathroom



Early Echoes

• Only one algorithm needed to simulate rooms of different sizes

• Group of delays (20) followed by a first order low pass filter

• Early echo output is the low pass filtered sum of each delay block



Early Echo Response - Auditorium



Unified System Response - Bathroom

• Combination of both early and late reflections• Uses medium room “tail” algorithm



Real-time Implementation

• Involves the rewriting of Gardner’s Network to process input signals on a sample-by-sample basis rather than processing bulk input samples (entire signals)

• Real-time implementation of non real-time functions: splitting up functions into two parts – real-time part & non real-time part

• Real-time implementation of the in-built Matlab filter function: two architectures – direct form I & II

• Trade-off – Network execution time vs. No. of memory elements needed



Translation from Matlab to C

• Real-time implementation on a suitable platform requires a translation of all real-time functions from Matlab to C

• C-code equivalent versions of all real-time Matlab functions were written and verified using the originals as references

• C test program results -> binary file -> read into Matlab -> verified against a Matlab equivalent program

• All results agreed perfectly!



C Program Execution

• Requires three user inputs:

• Sampling rate• Simulation file• Binary input file

name containing the samples to be reverberated



Functional Verification



Implementation on a DSP Development System

• ADSP-21065L EZ-KIT Lite evaluation kit from Analog Devices

• fast 32-bit DSP floating-point processor – 21065L

• a full duplex, 16-bit stereo audio codec – AD1819A (ADC & DAC)

• 544 Kb of internal user memory

• 16-megaword map of external SDRAM



Real-time Simulation of C-code on board

• Read in samples from a mic or CD player - ADC

• Process the samples in real-time using Gardner’s algorithm

• Send the samples out to a set of speakers - DAC

• Successful: Early Echo implementation

• Failure due to: lack of internal memory (big arrays), software tools…



Implementation on a PC-Based System

• 6025E DAQ plug-in device by National Instruments• A software programmable device (NI-DAQ

Functions) featuring 16 channels (eight differential) of analog input, two channels of analog output

• Two simulation programs:• ‘Pseudo’ real-time application – records n secs of

audio (mic/cd player), processes it, and plays it back in almost real-time

• Real-time application – records and plays back audio in real-time (low sampling rate < 2 kHz)



Conclusion

• Ultimate Project Goal: Real-time implementation of a reverberation algorithm for audio applications – practically realised

• A real-time audio reverberation system was implemented but only for low sampling rates

• A “pseudo” real-time system was also implemented• Future work: optimization of code to operate at

higher sampling rates (speech, audio)• Simulation of more room types other than a

bathroom, auditorium & cathedral

Implementation of an Audio Reverberation Algorithm Fergal Dunne Supervisor: Dr. Edward Jones.

Documents