vyonica-report 0.0.1

Objective voice-clone evaluation — metrics, graphs, and PDF reports. CLI + Python SDK.

vyonica-report

Objective voice-clone evaluation — CLI + Python SDK.

Computes ~50 objective metrics across acoustic quality, prosody, speaker identity, linguistic accuracy, and perceptual approximation. Emits:

• CSV (wide format) — stackable across many runs
• PNG graphs — waveform, pitch, energy, spectrogram, band energy
• PDF report — closeness score, verdict, all charts, metric evidence table

No LLM commentary. No database. No web server. Just metrics.

Install

pip install vyonica-report
# or, from a local checkout:
pip install /path/to/trainer/voice-cloning-result-comp/vyonica_report

Python 3.11+ required. The first run downloads ~2 GB of ML model weights (Whisper, ECAPA, NISQA, DNSMOS); subsequent runs are cached. Use --no-heavy to skip those models and keep the run pure-DSP (~5 s).

CLI

vyonica-report \
  --original ref.wav \
  --clone synth.wav \
  --transcript "The quick brown fox jumps over the lazy dog." \
  --language en \
  --output metrics.csv \
  --graphs graphs/ \
  --pdf report.pdf

Flag	Default	Purpose
--language	en	ISO 639-1 transcript language
--output / -o	metrics.csv	Wide-format CSV (one row per run)
--graphs DIR	—	Write diagnostic PNGs here
--pdf PATH	—	Write a PDF evaluation report here
--run-id ID	ISO timestamp	Label this row in the CSV
--append	off	Append a row instead of overwriting
--no-heavy	off	Skip Whisper / ECAPA / NISQA / DNSMOS / Praat
--whisper-model	small	Whisper size when heavy models are on

Python SDK

from vyonica_report import (
    analyze,
    to_csv,
    plot_graphs,
    generate_pdf,
    closeness_score,
)

metrics, summary = analyze(
    original_path="ref.wav",
    clone_path="synth.wav",
    transcript="The quick brown fox jumps over the lazy dog.",
    language="en",
    enable_heavy_models=True,
)

print(f"Voice closeness: {closeness_score(metrics):.1f}%")

to_csv(metrics, "metrics.csv", run_id="experiment_42")
plot_graphs(summary, "graphs/")
generate_pdf(metrics, summary, "report.pdf", run_label="experiment_42")

Each Metric has domain, name, value, unit, direction (higher / lower / near_1 / neutral), status, and note fields.

CSV format (wide)

One row per run; columns are {domain_slug}__{metric_slug}:

run_id,timestamp,acoustic_quality__log_spectral_distance,prosody__f0_contour_correlation,speaker_identity__mfcc_timbre_similarity,...
experiment_42,2026-05-23T11:22:33,2.31,0.87,0.91,...

Use --append to stack many runs into the same CSV.

Building & publishing

From the package directory:

# build wheel + sdist into dist/
uv build                   # or: python -m build

# install the built wheel into another local repo for testing
pip install dist/vyonica_report-0.1.0-py3-none-any.whl

# publish to PyPI (requires a token from https://pypi.org/manage/account/token/)
pip install twine
twine upload dist/*