ataiii commited on
Commit
688f7da
1 Parent(s): 1990dbb

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +131 -0
README.md ADDED
@@ -0,0 +1,131 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ tags:
3
+ - pyannote
4
+ - pyannote-audio
5
+ - pyannote-audio-pipeline
6
+ - audio
7
+ - voice
8
+ - speech
9
+ - speaker
10
+ - speaker-diarization
11
+ - speaker-change-detection
12
+ - voice-activity-detection
13
+ - overlapped-speech-detection
14
+ - automatic-speech-recognition
15
+ datasets:
16
+ - ami
17
+ - dihard
18
+ - voxconverse
19
+ - aishell
20
+ - repere
21
+ - voxceleb
22
+ license: mit
23
+ ---
24
+
25
+ # 🎹 Speaker diarization
26
+
27
+ Relies on pyannote.audio 2.0: see [installation instructions](https://github.com/pyannote/pyannote-audio/tree/develop#installation).
28
+
29
+
30
+ ## TL;DR
31
+
32
+ ```python
33
+ # load the pipeline from Hugginface Hub
34
+ from pyannote.audio import Pipeline
35
+ pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization@2022.07")
36
+
37
+ # apply the pipeline to an audio file
38
+ diarization = pipeline("audio.wav")
39
+
40
+ # dump the diarization output to disk using RTTM format
41
+ with open("audio.rttm", "w") as rttm:
42
+ diarization.write_rttm(rttm)
43
+ ```
44
+
45
+ ## Advanced usage
46
+
47
+ In case the number of speakers is known in advance, one can use the `num_speakers` option:
48
+
49
+ ```python
50
+ diarization = pipeline("audio.wav", num_speakers=2)
51
+ ```
52
+
53
+ One can also provide lower and/or upper bounds on the number of speakers using `min_speakers` and `max_speakers` options:
54
+
55
+ ```python
56
+ diarization = pipeline("audio.wav", min_speakers=2, max_speakers=5)
57
+ ```
58
+
59
+ If you feel adventurous, you can try and play with the various pipeline hyper-parameters.
60
+ For instance, one can use a more aggressive voice activity detection by increasing the value of `segmentation_onset` threshold:
61
+
62
+ ```python
63
+ hparams = pipeline.parameters(instantiated=True)
64
+ hparams["segmentation_onset"] += 0.1
65
+ pipeline.instantiate(hparams)
66
+ ```
67
+
68
+ ## Benchmark
69
+
70
+ ### Real-time factor
71
+
72
+ Real-time factor is around 5% using one Nvidia Tesla V100 SXM2 GPU (for the neural inference part) and one Intel Cascade Lake 6248 CPU (for the clustering part).
73
+
74
+ In other words, it takes approximately 3 minutes to process a one hour conversation.
75
+
76
+ ### Accuracy
77
+
78
+ This pipeline is benchmarked on a growing collection of datasets.
79
+
80
+ Processing is fully automatic:
81
+
82
+ * no manual voice activity detection (as is sometimes the case in the literature)
83
+ * no manual number of speakers (though it is possible to provide it to the pipeline)
84
+ * no fine-tuning of the internal models nor tuning of the pipeline hyper-parameters to each dataset
85
+
86
+ ... with the least forgiving diarization error rate (DER) setup (named *"Full"* in [this paper](https://doi.org/10.1016/j.csl.2021.101254)):
87
+
88
+ * no forgiveness collar
89
+ * evaluation of overlapped speech
90
+
91
+
92
+ | Benchmark | [DER%](. "Diarization error rate") | [FA%](. "False alarm rate") | [Miss%](. "Missed detection rate") | [Conf%](. "Speaker confusion rate") | Expected output | File-level evaluation |
93
+ | ---------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------- | --------------------------- | ---------------------------------- | ----------------------------------- | ------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------ |
94
+ | [AISHELL-4](http://www.openslr.org/111/) | 14.61 | 3.31 | 4.35 | 6.95 | [RTTM](reproducible_research/AISHELL.SpeakerDiarization.Full.test.rttm) | [eval](reproducible_research/AISHELL.SpeakerDiarization.Full.test.eval) |
95
+ | [AMI *Mix-Headset*](https://groups.inf.ed.ac.uk/ami/corpus/) [*only_words*](https://github.com/BUTSpeechFIT/AMI-diarization-setup) | 18.21 | 3.28 | 11.07 | 3.87 | [RTTM](reproducible_research/2022.07/AMI.SpeakerDiarization.only_words.test.rttm) | [eval](reproducible_research/2022.07/AMI.SpeakerDiarization.only_words.test.eval) |
96
+ | [AMI *Array1-01*](https://groups.inf.ed.ac.uk/ami/corpus/) [*only_words*](https://github.com/BUTSpeechFIT/AMI-diarization-setup) | 29.00 | 2.71 | 21.61 | 4.68 | [RTTM](reproducible_research/2022.07/AMI-SDM.SpeakerDiarization.only_words.test.rttm) | [eval](reproducible_research/2022.07/AMI-SDM.SpeakerDiarization.only_words.test.eval) |
97
+ | [CALLHOME](https://catalog.ldc.upenn.edu/LDC2001S97) [*Part2*](https://github.com/BUTSpeechFIT/CALLHOME_sublists/issues/1) | 30.24 | 3.71 | 16.86 | 9.66 | [RTTM](reproducible_research/2022.07/CALLHOME.SpeakerDiarization.CALLHOME.test.rttm) | [eval](reproducible_research/2022.07/CALLHOME.SpeakerDiarization.CALLHOME.test.eval) |
98
+ | [DIHARD 3 *Full*](https://arxiv.org/abs/2012.01477) | 20.99 | 4.25 | 10.74 | 6.00 | [RTTM](reproducible_research/2022.07/DIHARD.SpeakerDiarization.Full.test.rttm) | [eval](reproducible_research/2022.07/DIHARD.SpeakerDiarization.Full.test.eval) |
99
+ | [REPERE *Phase 2*](https://islrn.org/resources/360-758-359-485-0/) | 12.62 | 1.55 | 3.30 | 7.76 | [RTTM](reproducible_research/2022.07/REPERE.SpeakerDiarization.Full.test.rttm) | [eval](reproducible_research/2022.07/REPERE.SpeakerDiarization.Full.test.eval) |
100
+ | [VoxConverse *v0.0.2*](https://github.com/joonson/voxconverse) | 12.76 | 3.45 | 3.85 | 5.46 | [RTTM](reproducible_research/2022.07/VoxConverse.SpeakerDiarization.VoxConverse.test.rttm) | [eval](reproducible_research/2022.07/VoxConverse.SpeakerDiarization.VoxConverse.test.eval) |
101
+
102
+
103
+ ## Support
104
+
105
+ For commercial enquiries and scientific consulting, please contact [me](mailto:herve@niderb.fr).
106
+ For [technical questions](https://github.com/pyannote/pyannote-audio/discussions) and [bug reports](https://github.com/pyannote/pyannote-audio/issues), please check [pyannote.audio](https://github.com/pyannote/pyannote-audio) Github repository.
107
+
108
+
109
+ ## Citations
110
+
111
+ ```bibtex
112
+ @inproceedings{Bredin2021,
113
+ Title = {{End-to-end speaker segmentation for overlap-aware resegmentation}},
114
+ Author = {{Bredin}, Herv{\'e} and {Laurent}, Antoine},
115
+ Booktitle = {Proc. Interspeech 2021},
116
+ Address = {Brno, Czech Republic},
117
+ Month = {August},
118
+ Year = {2021},
119
+ }
120
+ ```
121
+
122
+ ```bibtex
123
+ @inproceedings{Bredin2020,
124
+ Title = {{pyannote.audio: neural building blocks for speaker diarization}},
125
+ Author = {{Bredin}, Herv{\'e} and {Yin}, Ruiqing and {Coria}, Juan Manuel and {Gelly}, Gregory and {Korshunov}, Pavel and {Lavechin}, Marvin and {Fustes}, Diego and {Titeux}, Hadrien and {Bouaziz}, Wassim and {Gill}, Marie-Philippe},
126
+ Booktitle = {ICASSP 2020, IEEE International Conference on Acoustics, Speech, and Signal Processing},
127
+ Address = {Barcelona, Spain},
128
+ Month = {May},
129
+ Year = {2020},
130
+ }
131
+ ```