zerobench_eval/README.md

zerobench_eval — the official ZeroBench-TTS scorer

Scores pre-generated wavs. It never loads, downloads, or runs a TTS model — you synthesize however you like, this reports the numbers.

pip install -r zerobench_eval/requirements.txt

python -m zerobench_eval manifest --out manifest.jsonl   # what to synthesize
#   ... your synthesis, one wav per row's `output_wav` ...
python -m zerobench_eval score --wav_dir my_wavs/ --name MyModel

Commands

command	what it does
manifest	writes one JSONL row per test item: text to say, ref_audio to clone, output_wav to write
score	scores a wav directory → per_sample.csv, summary.json, report.txt
rescore	recomputes WER from saved transcripts — no ASR, no GPU, runs in seconds

Layout

score looks for <wav_dir>/<subset>/<voice_id>.wav, and also accepts a nested wav/ folder or flat <subset>_<voice_id>.wav / <id>.wav names. If files are missing it tells you which and refuses to report a number, unless you pass --allow_missing (the summary is then flagged complete: false).

Files

file	contents
scorers.py	WER / SSIM / UTMOS / silence, self-contained
references.py	the acceptable-reference expansion — the core of the WER policy
benchmark.py	locating benchmark data, matching wavs to items
report.py	aggregation and the printed table
test_references.py	pins both directions of the WER policy — run it after any edit

Notes

• UTMOSv2 is optional. WER and SSIM work without it; pass --skip_utmos, or install it with pip install git+https://github.com/sarulab-speech/UTMOSv2.git.
• UTMOS is seeded. UTMOSv2 ensembles over random crops and is not reproducible unseeded (3.05 / 3.03 / 2.96 for the same clip). The RNG is reset before every clip so the score is a deterministic function of the audio.
• Don't change --asr if you want comparable numbers — the default pair is part of the benchmark definition.

Full metric definitions and the rationale are in the dataset README.