inclusionAI/Ming-Freeform-Audio-Edit-Benchmark

README

Introduction

This repository hosts Ming-Freeform-Audio-Edit, the benchmark test set for evaluating the downstream editing tasks of the Ming-UniAudio model.

This test set covers 7 distinct editing tasks, categorized as follows:

• Semantic Editing (3 tasks):

  • Free-form Deletion
  • Free-form Insertion
  • Free-form Substitution

• Acoustic Editing (5 tasks):

  • Time-stretching
  • Pitch Shifting
  • Dialect Conversion
  • Emotion Conversion
  • Volume Conversion

The audio samples are sourced from well-known open-source datasets, including seed-tts eval, LibriTTS, and Gigaspeech.

Dataset statistics

Semantic Editing

Task Types\ # samples \ Language	Zh deletion	Zh insertion	Zh substitution	En deletion	En insertion	En substitution
Index-based	186	180	36	138	100	67
Content-based	95	110	289	62	99	189
Total	281	290	325	200	199	256

Index-based instruction: specifies an operation on content at positions $i$ to $j$. (e.g. delete the characters or words from index 3 to 12)

Content-based: targets specific characters or words for editing. (e.g. insert 'hello' before 'world')

Acoustic Editing

Task Types\ # samples \ Language	Zh	En
Time-stretching	50	50
Pitch Shifting	50	50
Dialect Conversion	250	---
Emotion Conversion	84	72
Volume Conversion	50	50

Evaluation Metrics

Semantic Editing

For the deletion, insertion, and substitution tasks, we evaluate the performance using four key metrics:

• Word Error Rate (WER) of the Edited Region (wer)
• Word Error Rate (WER) of the Non-edited Region (wer.noedit)
• Edit Operation Accuracy (acc)
• Speaker Similarity (sim)

These metrics can be calculated by running the following command:

# run pip install -r requirements.txt first
bash eval_scripts/semantic/run_eval.sh /path/contains/edited/audios

NOTE: the directory passed to the above script should have the structure as follows:

.
├── del
│   └── edit_del_basic
│       ├── eval_result
│       ├── hyp.txt
│       ├── input_wavs
│       ├── origin_wavs
│       ├── ref.txt
│       ├── test.jsonl
│       ├── test_parse.jsonl # This is need to run the evaluation script
│       ├── test.yaml
│       └── tts/ # This is the directory contains the edited wavs

Examples of test_parse.jsonl:

{"uid": "00107947-00000092", "input_wav_path": "wavs/00107947-00000092.wav","output_wav_path": "edited_wavs/00107947-00000092.wav", "instruction": "Please recognize the language of this speech and transcribe it. And delete '随着经济的发'.\n", "asr_label": "随着经济的发展食物浪费也随之增长", "asr_text": "随着经济的发展食物浪费也随之增长", "edited_text_label": "展食物浪费也随之增长", "edited_text": "<edit></edit>展食物浪费也随之增长", "origin_speech_url": null,}

{"uid": "00010823-00000019", "input_wav_path": "wavs/00010823-00000019.wav", "output_wav_path": "edited_wavs/00010823-00000019.wav", "instruction": "Please recognize the language of this speech and transcribe it. And delete the characters or words from index 4 to index 10.\n", "asr_label": "我们将为全球城市的可持续发展贡献力量", "asr_text": "我们将为全球城市的可持续发展贡献力量", "edited_text_label": "我们将持续发展贡献力量", "edited_text": "我们将<edit></edit>持续发展贡献力量", "origin_speech_url": null}

Acoustic Editing

For the acoustic editing tasks, we use WER and SPK-SIM as the primary evaluation metrics. These two metrics can be calculated by running the following commands:

bash eval_scripts/acoustic/cal_wer_sim.sh /path/contains/edited/audios

Additionally, for the dialect and emotion conversion tasks, we assess the conversion accuracy by leveraging a large language model (LLM) through API calls.

# dialect conversion accuracy
python eval_scripts/acoustic/pyscripts/dialect_api.py --output_dir <保存评测结果的根目录> --generated_audio_dir <存放已生成音频文件的目录路径>
# emotion conversion accuracy
# fisrt, run: bash eval_scripts/acoustic/cal_wer_sim.sh /path/contains/edited/audios
python pyscripts/emo_acc.py