EVS

EVS Codec

This project provides the EVS(Enhanced Voice Services) codec optimized for Nuclei CPUs.

The source code comes from 3GPP TS 26.442 (current version :19.0.0).

Because EVS source code cannot be redistributed by this repository, only the patches and helper scripts are kept in-tree. The actual EVS source files must be downloaded from the official 3GPP archive and reconstructed locally before building.

File Structure

Directory	Description
data	test data and reference results
scripts	source download script and local patches
src	locally reconstructed EVS source tree after running the helper script
npk.tmp.yml	NPK template file, kept as a template because the repository does not ship EVS source code

Prerequests

Please refer to the Prerequests section in the parent directory's README.

Build

Before building, run the helper script below to download the official EVS archive, verify its checksum, extract the required source package, and apply the Nuclei-specific patches:

./scripts/download_and_patch.sh

scripts/download_and_patch.sh downloads the EVS release archive from 3GPP, checks the MD5 checksum, extracts the codec sources into EVS/src, applies rv32p_opt.patch and baremetal.patch, and removes files that are not needed for the bare-metal demo.

Since the repository only provides patches instead of redistributing EVS source code, the package description file is also kept as npk.tmp.yml rather than npk.yml. If you need an actual NPK package file, generate or rename it locally after reconstructing the source tree.

Switch to the directory containing the Makefile.
Below are build examples for the Nuclei N900 CPU.

Build without extension:

make CORE=n900 ARCH_EXT= all

Build with B and P extension:

make CORE=n900 ARCH_EXT=_zba_zbb_zbc_zbs_xxldspn1x all

For details about Nuclei CPU architecture extensions, see the ARCH_EXT section in the Nuclei SDK documentation.

Performance Test

Generate Test Data

Codec performance depends on both bitrate and input audio data.
To simplify testing with different bitrates and audio files, a helper script is provided:

👉 data/bin/gendata.sh

Note

This script is only supported on Linux x86_64
It relies on prebuilt binaries (encoder/decoder) available only for x86_64.
You may also build these binaries yourself if needed.

To view usage details:

./bin/gendata.sh -h

Expected Output

A typical output looks like this:

Nuclei SDK Build Time: Sep 29 2025, 09:40:40
Download Mode: DDR
CPU Frequency 50303139 Hz
CPU HartID: 0
Start Encoding...
Benchmark initialized

===========================================================================
 EVS Codec 3GPP TS26.442 August 12, 2021.
 Version 12.15.0 / 13.10.0 / 14.6.0 / 15.4.0 / 16.4.0
===========================================================================


Input audio file:       audio/stv48n2_1s.raw
Output bitstream file:  enc.192

Input sampling rate:    48000 Hz
Average bitrate:        5.90 kbps
DTX:                    ON, CNG update interval = 8 frames

Bandwidth limited to WB. To enable FB coding, please use -max_band FB.


------ Running the encoder ------
Frames processed:       50
Encoding finished
CSV, evs_encode, 117.84
Result matches!
Start Decoding...

===========================================================================
 EVS Codec 3GPP TS26.442 August 12, 2021.
 Version 12.15.0 / 13.10.0 / 14.6.0 / 15.4.0 / 16.4.0
===========================================================================


Input bitstream file:   enc.192
Output synthesis file:  dec.raw

Output sampling rate:   48000 Hz
Bitrate:                2.80 kbps

------ Running the decoder ------
Frames processed:       50
Decoding finished

CSV, evs_decode, 66.79
Result matches!
PASS

The key results are prefixed with CSV, making them easy to parse in scripts for further analysis.

The performance metric is reported in MCPS (Milion Cycles Per Second), which represents the number of CPU cycles required to process one second of audio data.