Update README.md
Browse files
README.md
CHANGED
@@ -24,13 +24,21 @@ The audio samples are sourced from well-known open-source datasets, including se
|
|
24 |
|
25 |
## Dataset statistics
|
26 |
### Semantic Editing
|
|
|
27 |
| Task Types\ # samples \ Language | Zh deletion | Zh insertion | Zh substitution | En deletion | En insertion | En substitution |
|
28 |
| -------------------------------- | ----------: | -----------: | --------------: | ----------: | -----------: | --------------: |
|
29 |
| Index-based | 186 | 180 | 36 | 138 | 100 | 67 |
|
30 |
| Content-based | 95 | 110 | 289 | 62 | 99 | 189 |
|
31 |
| Total | 281 | 290 | 325 | 200 | 199 | 256 |
|
32 |
|
33 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
|
35 |
*Content-based*: targets specific characters or words for editing. (e.g. insert 'hello' before 'world')
|
36 |
### Acoustic Editing
|
@@ -42,6 +50,13 @@ The audio samples are sourced from well-known open-source datasets, including se
|
|
42 |
| Emotion Conversion | 84 | 72 |
|
43 |
| Volume Conversion | 50 | 50 |
|
44 |
## Evaluation Metrics
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
### Semantic Editing
|
46 |
For the deletion, insertion, and substitution tasks, we evaluate the performance using four key metrics:
|
47 |
+ Word Error Rate (WER) of the Edited Region (wer)
|
@@ -49,44 +64,86 @@ For the deletion, insertion, and substitution tasks, we evaluate the performance
|
|
49 |
+ Edit Operation Accuracy (acc)
|
50 |
+ Speaker Similarity (sim)
|
51 |
|
52 |
-
|
53 |
-
```
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
│
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
|
73 |
-
Examples of test_parse.jsonl:
|
74 |
-
``` json
|
75 |
-
{"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,}
|
76 |
|
77 |
-
{"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}
|
78 |
-
```
|
79 |
### Acoustic Editing
|
80 |
-
For the acoustic editing tasks, we use WER and SPK-SIM as the primary evaluation metrics.
|
81 |
-
```bash
|
82 |
-
bash eval_scripts/acoustic/cal_wer_sim.sh /path/contains/edited/audios
|
83 |
-
```
|
84 |
|
85 |
-
|
86 |
-
```bash
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
|
|
|
|
|
|
|
|
|
|
|
24 |
|
25 |
## Dataset statistics
|
26 |
### Semantic Editing
|
27 |
+
#### full version
|
28 |
| Task Types\ # samples \ Language | Zh deletion | Zh insertion | Zh substitution | En deletion | En insertion | En substitution |
|
29 |
| -------------------------------- | ----------: | -----------: | --------------: | ----------: | -----------: | --------------: |
|
30 |
| Index-based | 186 | 180 | 36 | 138 | 100 | 67 |
|
31 |
| Content-based | 95 | 110 | 289 | 62 | 99 | 189 |
|
32 |
| Total | 281 | 290 | 325 | 200 | 199 | 256 |
|
33 |
|
34 |
+
#### basic version
|
35 |
+
| Task Types\ # samples \ Language | Zh deletion | Zh insertion | Zh substitution | En deletion | En insertion | En substitution |
|
36 |
+
| -------------------------------- | ----------: | -----------: | --------------: | ----------: | -----------: | --------------: |
|
37 |
+
| Index-based | 92 | 65 | 29 | 47 | 79 | 29 |
|
38 |
+
| Content-based | 78 | 105 | 130 | 133 | 81 | 150 |
|
39 |
+
| Total | 170 | 170 | 159 | 180 | 160 | 179 |
|
40 |
+
|
41 |
+
*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)
|
42 |
|
43 |
*Content-based*: targets specific characters or words for editing. (e.g. insert 'hello' before 'world')
|
44 |
### Acoustic Editing
|
|
|
50 |
| Emotion Conversion | 84 | 72 |
|
51 |
| Volume Conversion | 50 | 50 |
|
52 |
## Evaluation Metrics
|
53 |
+
### Environment Preparation
|
54 |
+
```bash
|
55 |
+
git clone https://github.com/inclusionAI/Ming-Freeform-Audio-Edit.git
|
56 |
+
cd Ming-Freeform-Audio-Edit
|
57 |
+
pip install -r requirements.txt
|
58 |
+
```
|
59 |
+
**Note**: Please download the audio and meta files from [HuggingFace](https://huggingface.co/datasets/inclusionAI/Ming-Freeform-Audio-Edit-Benchmark/tree/main) or [ModelScope](https://modelscope.cn/datasets/inclusionAI/Ming-Freeform-Audio-Edit-Benchmark/files) and put the `wavs` and `meta` directories under `Ming-Freeform-Audio-Edit`
|
60 |
### Semantic Editing
|
61 |
For the deletion, insertion, and substitution tasks, we evaluate the performance using four key metrics:
|
62 |
+ Word Error Rate (WER) of the Edited Region (wer)
|
|
|
64 |
+ Edit Operation Accuracy (acc)
|
65 |
+ Speaker Similarity (sim)
|
66 |
|
67 |
+
1. If you have organized the directories contain edited waveforms like below:
|
68 |
+
```
|
69 |
+
eval_path
|
70 |
+
|
|
71 |
+
├── del
|
72 |
+
│ └── edit_del_basic
|
73 |
+
│ └── tts/ # This is the actual directory contains the edited wavs
|
74 |
+
├── ins
|
75 |
+
│ └── edit_ins_basic
|
76 |
+
│ └── tts/ # This is the actual directory contains the edited wavs
|
77 |
+
├── sub
|
78 |
+
└── edit_sub_basic
|
79 |
+
└── tts/ # This is the actual directory contains the edited wavs
|
80 |
+
```
|
81 |
+
Then you can run the following command to get those metrics:
|
82 |
+
```bash
|
83 |
+
cd Ming-Freeform-Audio-Edit/eval_scripts
|
84 |
+
bash run_eval_semantic.sh eval_path \
|
85 |
+
whisper_path \
|
86 |
+
paraformer_path \
|
87 |
+
wavlm_path \
|
88 |
+
eval_mode \
|
89 |
+
lang
|
90 |
+
```
|
91 |
+
Here is a brief description of the parameters for the script above:
|
92 |
+
+ `eval_path`: The top-level directory containing subdirectories for each editing task
|
93 |
+
+ `whisper_path`:Path to the Whisper model, which is used to calculate WER for English audio. You can download it from [here](https://huggingface.co/openai/whisper-large-v3).
|
94 |
+
+ `paraformer_path`:Path to the Paraformer model, which is used to calculate WER for Chinese audio. You can download it from [here](https://huggingface.co/funasr/paraformer-zh).
|
95 |
+
+ `wavlm_path`: Path to the WavLM model, which is used to calculate speaker similarity. You can download it from [here](https://drive.google.com/file/d/1-aE1NfzpRCLxA4GUxX9ITI3F9LlbtEGP/view).
|
96 |
+
+ `eval_mode`: Used to specify which version of the evaluation set to use. Choose between `basic` and `open`
|
97 |
+
+ `lang`: supported language, choose between `zh` and `en`
|
98 |
+
|
99 |
+
2. If your directory for the edited audio is not organized in the format described above, you can run the following commands.
|
100 |
+
```bash
|
101 |
+
cd eval_scripts
|
102 |
+
# get wer, wer.noedit
|
103 |
+
bash cal_wer_edit.sh meta_file \
|
104 |
+
wav_dir \
|
105 |
+
lang \
|
106 |
+
num_jobs \
|
107 |
+
res_dir \
|
108 |
+
task_type \
|
109 |
+
eval_mode \
|
110 |
+
whisper_path \
|
111 |
+
paraformer_path \
|
112 |
+
edit_cat # use `semantic` here
|
113 |
+
# get sim
|
114 |
+
bash cal_sim_edit.sh meta_file \
|
115 |
+
wav_dir \
|
116 |
+
wavlm_path \
|
117 |
+
num_jobs \
|
118 |
+
res_dir \
|
119 |
+
lang
|
120 |
+
```
|
121 |
+
Here is a brief description of the parameters for the script above:
|
122 |
+
+ `meta_file`: The absolute path to the meta file for the corresponding task (e.g., `meta_en_deletion_basic.csv` or `meta_en_deletion.csv`).
|
123 |
+
+ `wav_dir`: The directory containing the edited audio files (the WAV files should be located directly in this directory).
|
124 |
+
+ `lang`: `zh` or `en`
|
125 |
+
+ `num_jobs`: number of process.
|
126 |
+
+ `res_dir`: The directory to save the metric results.
|
127 |
+
+ `task_type`: `del`, `ins` or `sub`
|
128 |
+
+ `eval_mode`: The same as the above.
|
129 |
+
+ `whisper_path`: The same as the above
|
130 |
+
+ `paraformer_path`: The same as the above
|
131 |
+
+ `edit_cat`: `semantic` or `acoustic`
|
132 |
|
|
|
|
|
|
|
133 |
|
|
|
|
|
134 |
### Acoustic Editing
|
135 |
+
For the acoustic editing tasks, we use WER and SPK-SIM as the primary evaluation metrics.
|
|
|
|
|
|
|
136 |
|
137 |
+
1. If the directory for the edited audio is structured, you can run the following command.
|
138 |
+
```bash
|
139 |
+
cd Ming-Freeform-Audio-Edit/eval_scripts
|
140 |
+
bash run_eval_acoustic.sh eval_path \
|
141 |
+
whisper_path \
|
142 |
+
paraformer_path \
|
143 |
+
wavlm_path \
|
144 |
+
eval_mode \
|
145 |
+
lang
|
146 |
+
```
|
147 |
+
2. Otherwise, you can run commands similar to the one for the semantic tasks, with the `edit_cat` parameter set to `acoustic`.
|
148 |
+
|
149 |
+
Additionally, for the dialect and emotion conversion tasks, we assess the conversion accuracy by leveraging a large language model (LLM) through API calls, refer to `eval_scripts/run_eval_acoustic.sh` for more details.
|