Skip to content
Murali RaviMedical Instrumentation
Proof of concept2025-2026

Edge-Deployable Voice Biometric Authentication

Speaker verification at 2.17% EER, compressed to 1.57 MiB, running on the device itself.

Developed in collaboration with Analog Devices India, this project targets high-accuracy speaker verification on resource-constrained edge hardware: specifically the ARM Cortex-A55. The constraint is unforgiving: the entire pipeline, from feature extraction to verification decision, has to fit and run on-chip.

0.00% EER

on the VoxCeleb1 test set

Equal Error Rate (float)
2.17%
EER after INT8 quantisation
2.60%
Quantised model size
1.57 MiB
Channels (ECAPA-Lite)
128
ECAPA-LITE: 128-CHANNEL LIGHTWEIGHT ECAPA-TDNN VARIANTWaveform16 kHz audioMFCCOn-chipConv1D128 chSE-Res2Block×3, dilatedAttentive StatPoolingEmbeddingSpeaker vectorCHANNEL REDUCTION: WHY LITEECAPA-TDNN512 channelsECAPA-LITE128 channelsSELECTED FOR ACCURACY-TO-SIZE RATIO: THE METRIC THAT MATTERS WHEN DEPLOYMENT IS MEASURED IN MEGABYTESEQUAL ERROR RATE2.17%VoxCeleb1 test set, full precisionCHANNELS128Lightweight variant widthQUANTISED SIZE1.57 MiBINT8 post-training

Summary

What the work set out to do, and the reasoning that shaped it.

The system is built on ECAPA-Lite, a 128-channel lightweight variant of the ECAPA-TDNN architecture, selected for its exceptional accuracy-to-size ratio: the property that matters most when the deployment target is measured in megabytes.

A significant innovation of this work is the addition of a Lightweight Prosody Encoder branch. This captures supra-segmental features: pitch, energy and duration: which significantly enhance the system's ability to discriminate between speakers. Prosody carries identity information that spectral features alone leave on the table.

The model achieved an Equal Error Rate of 2.17% on the VoxCeleb1 test set. To meet edge deployment requirements it was quantised to INT8 precision, reducing size to approximately 1.57 MiB while maintaining high accuracy at 2.60% EER: a 0.43 percentage point cost for a substantial compression gain.

Impact

What changed as a result.

01

Demonstrates that speaker verification competitive on a standard public benchmark can run entirely on edge hardware, with no network round-trip and no server-side inference.

02

The prosody encoder branch is a novel contribution: supra-segmental features are added at negligible parameter cost while measurably improving speaker discrimination.

03

The complete solution includes on-chip feature extraction: both MFCC and prosody: and was validated on the EV-SC598-SOM platform, proving feasibility for low-power, real-world biometric security.

Figures

The work, drawn.

Purpose-built diagrams: each one carries an idea that prose alone would take a page to deliver.

ECAPA-LITE: 128-CHANNEL LIGHTWEIGHT ECAPA-TDNN VARIANTWaveform16 kHz audioMFCCOn-chipConv1D128 chSE-Res2Block×3, dilatedAttentive StatPoolingEmbeddingSpeaker vectorCHANNEL REDUCTION: WHY LITEECAPA-TDNN512 channelsECAPA-LITE128 channelsSELECTED FOR ACCURACY-TO-SIZE RATIO: THE METRIC THAT MATTERS WHEN DEPLOYMENT IS MEASURED IN MEGABYTESEQUAL ERROR RATE2.17%VoxCeleb1 test set, full precisionCHANNELS128Lightweight variant widthQUANTISED SIZE1.57 MiBINT8 post-training
01

ECAPA-Lite architecture

A 128-channel lightweight ECAPA-TDNN variant chosen for accuracy-to-size ratio rather than headline accuracy alone.

LIGHTWEIGHT PROSODY ENCODER: A PARALLEL BRANCH FOR SUPRA-SEGMENTAL FEATURESINPUTWaveformBRANCH 01: SPECTRALMFCC → ECAPA-LiteVocal tract shape · segmental detailBRANCH 02: PROSODY · NOVELLightweight EncoderPitchF0 contourEnergyIntensity envelopeDurationTiming & rhythmFUSIONSpeaker embeddingEnhanced discriminationbetween speakersWHY IT ADDS RATHER THAN DUPLICATESProsody varies slowly and encodes speaking style; MFCCs capture vocal tract shape.THE TWO CARRY DIFFERENT IDENTITY INFORMATION: WHICH IS WHY THE BRANCH IMPROVES DISCRIMINATION AT NEGLIGIBLE PARAMETER COST.
02

The prosody encoder branch

Pitch, energy and duration run in parallel with the spectral path. Prosody encodes speaking style: information MFCCs discard.

INT8 QUANTISATION: WHAT IT COSTS, WHAT IT BUYSEQUAL ERROR RATE · VOXCELEB1Float precision2.17%INT8 quantised2.6%ACCURACY COST+0.43 ppMODEL FOOTPRINTFLOAT32INT8DEPLOYED SIZE≈ 1.57 MiBTHE TRADE0.43 percentage points of accuracy, exchanged for a model that fits on the device.A CONTROLLED AND ACCEPTABLE COST FOR THE EDGE DEPLOYMENT ENVELOPE.
03

The compression trade

INT8 quantisation costs 0.43 percentage points of EER and buys a model that fits in 1.57 MiB.

ON-DEVICE PIPELINE: EV-SC598-SOM · ARM CORTEX-A55EV-SC598-SOMPROCESSORARMCortex-A55RESOURCE-CONSTRAINEDMic inputAudio captureMFCCOn-chip extractionProsodyOn-chip extractionECAPA-LiteINT8 inferenceVerifyScore vs enrolmentNETWORK BOUNDARYNo round-tripNO SERVER INFERENCENO AUDIO LEAVES DEVICEAPPLICATIONIn-vehicle voice authentication: Analog Devices, Inc. proof of concept, 2025-2026.
04

On-device pipeline

Feature extraction, embedding and verification all execute on the ARM Cortex-A55. Nothing leaves the device.

Methods

How it was done.

ECAPA-Lite backbone

A 128-channel lightweight variant of ECAPA-TDNN. The channel reduction is what brings the parameter count into edge range while preserving the architecture's attentive statistics pooling.

Lightweight Prosody Encoder

A parallel branch capturing supra-segmental features: pitch, energy, duration. These vary slowly relative to spectral content and encode speaking style rather than vocal tract shape, which is why they add discriminative power rather than duplicating it.

INT8 quantisation

Post-training quantisation to INT8 precision reduces the model to approximately 1.57 MiB. Accuracy degrades from 2.17% to 2.60% EER: a controlled and acceptable trade for the deployment envelope.

On-chip feature extraction

MFCC and prosody feature extraction both run on-device, so no part of the pipeline depends on off-chip preprocessing.

Hardware validation

Validated on the EV-SC598-SOM platform targeting the ARM Cortex-A55, establishing feasibility rather than merely simulating it.

Specifications

The numbers.

Architecture
ECAPA-Lite128-channel lightweight ECAPA-TDNN variant
Novel component
Lightweight Prosody Encoder branch
Prosodic features
Pitch, energy, duration
Spectral features
MFCCExtracted on-chip
Benchmark
VoxCeleb1 test set
EER (float)
2.17%
EER (INT8)
2.60%
Quantised size
≈ 1.57 MiB
Target processor
ARM Cortex-A55
Validation platform
EV-SC598-SOM

Development timeline

From simulation to the clinic.

2025

Project sanctioned

Analog Devices, Inc. awards ₹6 Lakhs for a proof of concept in voice authentication for cars using an ML approach. Dr. Ravi serves as Principal Investigator.

Architecture

ECAPA-Lite selection

A 128-channel lightweight ECAPA-TDNN variant chosen for its accuracy-to-size ratio under edge constraints.

Innovation

Prosody encoder branch

A lightweight parallel branch added to capture pitch, energy and duration: supra-segmental cues that improve speaker discrimination.

Benchmark

2.17% EER on VoxCeleb1

Full-precision model evaluated on the VoxCeleb1 test set.

Compression

INT8 quantisation

Model reduced to ≈1.57 MiB at 2.60% EER, meeting the edge deployment envelope.

2026

On-device validation

Complete pipeline with on-chip MFCC and prosody extraction validated on EV-SC598-SOM.

Outcomes

  • ECAPA-Lite with a novel Lightweight Prosody Encoder branch
  • 2.17% EER on VoxCeleb1 at full precision
  • 2.60% EER at INT8, ≈1.57 MiB model size
  • On-chip MFCC and prosody feature extraction
  • Validated on EV-SC598-SOM targeting ARM Cortex-A55

Collaborators & support

  • Analog Devices India
  • Sri Sathya Sai Institute of Higher Learning

Funded under

Analog Devices, Inc. · INR 6 Lakhs

2025-2026 · Principal Investigator