ViSQOL-based quality and compatibility tests for Codec2-mod vs. vanilla Codec2 (3200 mode), including cross-encoding between the two.
- gcc, make, git
- libsndfile and libcodec2 (development packages)
- sox
- ViSQOL (
visqolinPATH, speech-mode lattice model) - A
wav_files/directory with mono, 8 kHz, 16-bit PCM WAV files
The ViSQOL model path can be changed with VISQOL_MODEL=/path/to/model.tflite.
make # clone/pull Codec2-mod branches and build
make run # default: main branch, 20 files
make run MOD=split-no-doubles N=all # use a different branch, all files
make run N=50 SEED=3 JOBS=4 # 50 files, selection seed 3, 4 parallel jobs
make always pulls the latest version of each branch. Files are processed in parallel, one per CPU core by default. It is recommended, as ViSQOL is CPU-heavy.
Each file is encoded with both codecs and decoded with both, and every combination is scored against the original with ViSQOL (resampled to 16 kHz).
| Column | Meaning |
|---|---|
ref-ref, mod-mod, ref-mod, mod-ref |
MOS for encoder-decoder pairs (ref-mod = vanilla encoder, Codec2-mod decoder) |
dMOS |
mod-mod − ref-ref: total change |
dMOS-dec |
ref-mod − ref-ref: decoder effect (same vanilla bits) |
dMOS-enc |
mod-ref − ref-ref: encoder effect (same vanilla decoder) |
Bit-diff |
Mismatched bits between the two bitstreams (- = identical) |
All deltas are relative to vanilla Codec2: green = better, red = worse. Per-file results are saved to work/results.csv, and errors to work/errors.log.
make clean # clean up binaries only
make distclean # also clean test results and Codec2-mod checkouts
