guynich commited on
Commit
62e2c8e
·
verified ·
1 Parent(s): c73a9c8

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +13 -18
README.md CHANGED
@@ -13,16 +13,11 @@ Voice Activity Detection (VAD) Test Dataset
13
 
14
  This dataset is based on the `test.clean` and `test.other` splits from the
15
  [librispeech_asr](https://huggingface.co/datasets/openslr/librispeech_asr)
16
- corpus. It includes two binary features:
17
 
18
  - **speech**: Indicates presence of speech ([0, 1]), computed using a dynamic threshold method with background noise estimation and smoothing.
19
 
20
- - **confidence**: A post-processing flag to optionally correct transient dropouts in speech. It is set to 1 by default, but switches to 0 for up to ~0.1 seconds (3 chunks of audio) following a transition from speech to silence. Approximately 7% of the `speech` features in this dataset are `confidence` 0. The remaining 93% are `confidence` 1 enabling VAD testing.
21
-
22
- | split | active speech, hours | confidence, % |
23
- | ---------- | :------------------: | :-----------: |
24
- | test.clean | 5.4 | 93.2 |
25
- | test.other | 5.3 | 92.6 |
26
 
27
  The dataset has minimal background noise, making it suitable for mixing with
28
  external noise samples to test VAD robustness.
@@ -66,17 +61,20 @@ In practice you would run the AUC computation across the entire test split.
66
 
67
  ## Ignore transient dropouts
68
 
69
- Filtering out low-confidence frames (~7% of the data) improves precision and
70
- reduces false positives, helping account for slower VAD responses in real-world
71
- applications.
 
72
 
73
  ```console
74
  confidence = dataset["test.clean"][0]["confidence"]
75
- confidence_mask = (np.array(confidence) == 1)
 
 
76
 
77
  roc_auc_confidence = roc_auc_score(
78
- np.array(speech)[confidence_mask],
79
- np.array(speech_probs)[confidence_mask],
80
  )
81
  ```
82
 
@@ -89,14 +87,11 @@ model with `test.clean` split.
89
  <img src="assets/roc_test_clean.png" alt="Example from test.clean with Silero-VAD"/>
90
 
91
  Precision values are increased when data is sliced by `confidence` values.
92
- These low-confidence `speech` features are flagged rather than removed, allowing
93
  users to either exclude them (as shown here) or handle them with other methods.
94
 
95
  <img src="assets/roc_test_clean_exclude_low_confidence.png" alt="Example from test.clean with Silero-VAD"/>
96
 
97
- Code to generate these plots is available on
98
- [GitHub](https://github.com/guynich/vad_eval_curves).
99
-
100
  # License Information
101
 
102
  This derivative dataset retains the same license as the source dataset
@@ -107,7 +102,7 @@ This derivative dataset retains the same license as the source dataset
107
 
108
  # Citation Information
109
 
110
- Features contributed by Guy Nicholson were added to the following dataset.
111
  ```
112
  @inproceedings{panayotov2015librispeech,
113
  title={Librispeech: an ASR corpus based on public domain audio books},
 
13
 
14
  This dataset is based on the `test.clean` and `test.other` splits from the
15
  [librispeech_asr](https://huggingface.co/datasets/openslr/librispeech_asr)
16
+ corpus. It includes two binary labels:
17
 
18
  - **speech**: Indicates presence of speech ([0, 1]), computed using a dynamic threshold method with background noise estimation and smoothing.
19
 
20
+ - **confidence**: A post-processing flag to optionally correct transient dropouts in speech. It is set to 1 by default, but switches to 0 for up to ~0.1 seconds (3 chunks of audio) following a transition from speech to silence. Approximately 7% of the `speech` labels in this dataset are `confidence` 0. The remaining 93% are `confidence` 1 enabling VAD testing.
 
 
 
 
 
21
 
22
  The dataset has minimal background noise, making it suitable for mixing with
23
  external noise samples to test VAD robustness.
 
61
 
62
  ## Ignore transient dropouts
63
 
64
+ The `confidence` values can be used to filter the data. Removing zero confidence
65
+ values excludes 6.8% of the dataset and causes numerical increase in
66
+ computed precision. This compensates for slower moving voice activity decisions
67
+ as encountered in real-world applications.
68
 
69
  ```console
70
  confidence = dataset["test.clean"][0]["confidence"]
71
+
72
+ speech_array = np.array(speech)
73
+ speech_probs_array = np.array(speech_probs)
74
 
75
  roc_auc_confidence = roc_auc_score(
76
+ speech_array[np.array(confidence) == 1],
77
+ speech_probs_array[np.array(confidence) == 1],
78
  )
79
  ```
80
 
 
87
  <img src="assets/roc_test_clean.png" alt="Example from test.clean with Silero-VAD"/>
88
 
89
  Precision values are increased when data is sliced by `confidence` values.
90
+ These low-confidence `speech` labels are flagged rather than removed, allowing
91
  users to either exclude them (as shown here) or handle them with other methods.
92
 
93
  <img src="assets/roc_test_clean_exclude_low_confidence.png" alt="Example from test.clean with Silero-VAD"/>
94
 
 
 
 
95
  # License Information
96
 
97
  This derivative dataset retains the same license as the source dataset
 
102
 
103
  # Citation Information
104
 
105
+ Labels contributed by Guy Nicholson were added to the following dataset.
106
  ```
107
  @inproceedings{panayotov2015librispeech,
108
  title={Librispeech: an ASR corpus based on public domain audio books},