Skip to content

Latest commit

 

History

History
86 lines (59 loc) · 2.8 KB

applying.md

File metadata and controls

86 lines (59 loc) · 2.8 KB
Raw

Things to check before applying

Make sure your recordings are under the .wav format and have been sampled at 16 kHz, to do so you can use :

soxi /path/to/my_recording.wav

You also need to make sure that you've activated the conda environment you created :

conda activate pyannote # or source activate pyannote, depending on your config

Last but not least, make sure that you are in the voice_type_classifier repository.

Applying

The code consists mainly of a wrapper of pyannote-audio's code responsible for applying a model. If you want to go further by retraining the model, you should go around here.

You can apply our pre-trained model by typing :

./apply.sh /path/to/my_recordings/ --device=gpu

where /path/to/my_recordings/ is a folder containing audio recordings in the .wav format (single-channel, 16kHz). This will actually apply 5 models (each for one of the class), which can take some time. The flag --device=gpu indicates if the model should be run on gpu or not. If this flag is not provided, the model will run on CPU.

You can also apply the model to a single audio file :

./apply.sh /path/to/my_recordings/my_recording.wav --device=gpu

Output format

The above commands will generate a output_voice_type_classifier folder, itself containing sub-folders named after the folder you applied the model on. Let's assume we applied the model on the my_recordings1 and my_recordings2 folder :

ls output_voice_type_classifier/

returns

my_recordings1  my_recordings2

Now, let's have a look at what my_recordings1 contains :

ls output_voice_type_classifier/my_recordings1

returns

CHI.rttm  FEM.rttm  KCHI.rttm  MAL.rttm  SPEECH.rttm  all.rttm

Each of the rttm files contains lines that look like :

SPEAKER	my_recording_number1	1	33.429	0.513	<NA>	<NA>	KCHI	<NA>	<NA>
SPEAKER	my_recording_number1	1	35.785	1.117	<NA>	<NA>	KCHI	<NA>	<NA>
SPEAKER	my_recording_number2	1	0.107	0.154	<NA>	<NA>	KCHI	<NA>	<NA>
SPEAKER	my_recording_number2	1	1.284	0.332	<NA>	<NA>	KCHI	<NA>	<NA>

Where :

  • each line corresponds to a speech/vocalization utterance
  • the second column is the name of the file concerned by this utterance
  • the fourth column is the onset of the utterance
  • the fifth column is the duration of the utterance
  • the eigth column is the voice type, amongst [KCHI, CHI, MAL, FEM, SPEECH]

CHI.rttm contains only utterances produced by other children (according to the model).

MAL.rttm contains only utterances produced by male speakers (according to the model)

And so on ...

all.rttm is the concatenation (sorted by filename and onset) of {KCHI,CHI,MAL,FEM,SPEECH}.rttm