Speech Quality Subjective Evaluation

This repository provides four general subjective tests (MOS: mean opinion score; PK: preference test; SIM: speaker similarity test; XAB: XAB test) for voice conversion(VC) and speech synthesis (SS) evaluations. The evaluation results will output in an Excel file with statistic results.

Subjective evaluations

For VC/SS, subjective evaluations are usually conducted. The general performance measurements of VC/SS are speech quality (the naturalness of converted/synthesized speech) and speaker similarity (the similarity between converted speech and target speech). The repository includes four types of subjective tests for VC/SS performance evaluation. Each test plays several speech files to listeners and ask them to give a specific score to evaluate each converted/synthesized file. The system will output an average score with a confidence interval of each test into a Excel file.

Mean opinion score (MOS):

• Speech quality (naturalness) evaluation
• System plays a speech file to a listener and asks him/her to give a score (1-5) to evaluate the speech quality of this speech file.

Speaker similarity test (SIM):

• Speaker similarity evaluation
• Proposed by the Voice Conversion Challenge (VCC) organize
• System plays a golden and a predicted speech files to a listener and asks him/her to measure that these two speeches are from the same speaker or not.
• Four measurements are given: 1. Definitely the same; 2. Maybe the same; 3. Maybe different; 4. Definitely different.

Preference test (PK):

• Speech quality or speaker similarity evaluations
• System plays two speech files of different methods to a listener and asks him/her to pick up the file with better performance.

XAB test:

• Speaker similarity or speech quality evaluation
• System plays a golden file and two speech files of different methods to a listener and ask him/her to pick up the file that is more similar to the golden speech.

---

Setup

Install requirements

pip install -r requirements.txt

# For Mac user
pip install -U PyObjC 

Folder structure

• data/{project}/: testing data of {project}.
• data/example/: example speech files of XAB and MOS tests.
• results: output Excel file.
• src: source code.
• yml: testing configs and profiles.

---

Usage

• Take project EVAL1 as an example.

1. Data preparation

• Create data/EVAL1/ folder.
• Put testing data of each method in data/EVAL1/{method}.
• Create data lists of all methods and put them in the same folder.

2. Config initialization

• Create yml/EVAL1/ folder
• Create evaluation.yml, record.yml, test_system.yml, and text.yml in yml/EVAL1/ folder.

3. Parse testing profile

• Create .yml format testing profile files for each testing type and subset corresponding to the evaluation.yml config.

python run_pre.py EVAL1

4. Run evaluation

• Run evaluation and the results will be in the /result/EVAL1_{type}.xlsx files. Each listener's results will be in the yml/EVAL1/{type}_{subset}_{userID}.yml files.
• -o {mode}: set mode to control the output information; choice:[vc, as, others].
• Set mode 'vc' will output complete details and mode 'others' will only output overall results.

python run_test.py EVAL1 -o {mode}

5. Get statistics

• Output evaluation results with statistics to the /result/EVAL1_final_{type}.xlsx files.

python run_post.py EVAL1 -o {mode}

---

COPYRIGHT

Copyright (c) 2020 Yi-Chiao Wu @ Nagoya University (@bigpon)
E-mail: yichiao.wu@g.sp.m.is.nagoya-u.ac.jp
Released under Apache 2.0