pyannote / pyannote-audio Goto Github PK

Hi, I am trying to accelerate the training process with my GPU, only to find a high GPU memory usage while a ZERO GPU-Util, as shown in the following picture.

Is there any appropriate way to adopt GPU during training of triplet models?

Ps. Both of these two threads are for triplet loss training. No other application are adopting GPU.

• Links: a high-dimensional online clustering method
• Information bottleneck
• Clustering Millions of Faces by Identity
• SpectralNet: Spectral Clustering using Deep Neural Networks
• von Mises-Fisher clustering
  See "Leveraging Mixture Structure from Multivariate von Mises-Fisher Distributions for Informed Speaker Clustering for Naturalistic Audio Streams" at Interspeech 2018
• See parsing approaches...
• Speaker clustering using dominant set
• UMAP + HDBSCAN
• ClusterNet
• Sampling clustering
• Spectral clustering used in "Speaker diarization with LSTM"
• Joint Unsupervised Learning of Deep Representations and Image Clusters (code, paper)
• spherecluster python library provides implementation of spherical K-means and von Mises-Fisher clustering

Tools

• pyroomacoustics Pyroomacoustics is a package for audio signal processing for indoor applications. It was developed as a fast prototyping platform for beamforming algorithms in indoor scenarios.
• AudioTSM is a python library for real-time audio time-scale modification procedures, i.e. algorithms that change the speed of an audio signal without changing its pitch.
• CLEESE is a sound-manipulation software tool designed to generate an infinite number of natural-sounding, expressive variations around an original speech recording.
• pydub, manipulate audio with a simple and easy high level interface
• pysox Python wrapper around sox
• Data Augmentation For Wearable Sensor Data
• muda: A library for augmenting annotated audio data
• maracas is a library for corrupting audio files with additive and convolutive noise.
• pySpeechRev efficient speech reverberation starting from a dataset of close-talking speech signals and a collection of acoustic impulse responses.
• Reverb class in PASE
• rir_simulator_python

Databases

• DEMAND

When running the apply-command I get a following error message. This comes after I have run the tune-command with the result that I describe in issue #39.

/Users/niko/anaconda3/envs/py35-pyannote-audio/lib/python3.5/site-packages/keras/models.py:245: UserWarning: No training configuration found in save file: the model was *not* compiled. Compile it manually.
  warnings.warn('No training configuration found in save file: '
Test set: 0it [00:00, ?it/s]
Traceback (most recent call last):
  File "/Users/niko/anaconda3/envs/py35-pyannote-audio/bin/pyannote-speech-detection", line 11, in <module>
    sys.exit(main())
  File "/Users/niko/anaconda3/envs/py35-pyannote-audio/lib/python3.5/site-packages/pyannote/audio/applications/speech_detection.py", line 615, in main
    application.apply(protocol_name, subset=subset)
  File "/Users/niko/anaconda3/envs/py35-pyannote-audio/lib/python3.5/site-packages/pyannote/audio/applications/speech_detection.py", line 565, in apply
precomputed = Precomputed(root_dir=apply_dir)
  File "/Users/niko/anaconda3/envs/py35-pyannote-audio/lib/python3.5/site-packages/pyannote/audio/features/utils.py", line 110, in __init__
    start = f.attrs['start']
  File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper (/Users/travis/miniconda3/conda-bld/h5py_1496412421941/work/h5py-2.7.0/h5py/_objects.c:2846)
  File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper (/Users/travis/miniconda3/conda-bld/h5py_1496412421941/work/h5py-2.7.0/h5py/_objects.c:2804)
  File "/Users/niko/anaconda3/envs/py35-pyannote-audio/lib/python3.5/site-packages/h5py/_hl/attrs.py", line 58, in __getitem__
    attr = h5a.open(self._id, self._e(name))
  File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper (/Users/travis/miniconda3/conda-bld/h5py_1496412421941/work/h5py-2.7.0/h5py/_objects.c:2846)
  File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper (/Users/travis/miniconda3/conda-bld/h5py_1496412421941/work/h5py-2.7.0/h5py/_objects.c:2804)
  File "h5py/h5a.pyx", line 77, in h5py.h5a.open (/Users/travis/miniconda3/conda-bld/h5py_1496412421941/work/h5py-2.7.0/h5py/h5a.c:2343)
KeyError: "Can't open attribute (Can't locate attribute: 'start')"

Hi Bredin!
I was trying to replicate feature-extraction tutorial. I followed the installation instructions as given in master branch.
But when I tried to execute command cat ~/.pyannote/db.yml, I got the error as follows:

cat: /home/abhishek/.pyannote/db.yml: No such file or directory

So, after skipping this, when I tried to extract the features from audio files pyannote-speech-feature ${EXPERIMENT_DIR} GameOfThrones I got the following error:

/home/abhishek/anaconda3/envs/pyannote-audio/lib/python3.5/site-packages/h5py/init.py:36: FutureWarning: Conversion of the second argument of issubdtype from float to np.floating is deprecated. In future, it will be treated as np.float64 == np.dtype(float).type.
from ._conv import register_converters as _register_converters
Traceback (most recent call last):
File "/home/abhishek/anaconda3/envs/pyannote-audio/bin/pyannote-speech-feature", line 11, in
sys.exit(main())
File "/home/abhishek/anaconda3/envs/pyannote-audio/lib/python3.5/site-packages/pyannote/audio/applications/feature_extraction.py", line 188, in main
preprocessors = {'wav': FileFinder(db_yml)}
File "/home/abhishek/anaconda3/envs/pyannote-audio/lib/python3.5/site-packages/pyannote/database/util.py", line 85, in init
with open(config_yml, 'r') as fp:
FileNotFoundError: [Errno 2] No such file or directory: '/home/abhishek/.pyannote/db.yml'

So, I have the following doubts:

1. What is db.yml file?
2. How is this file created? I created the pyannote.database plugin for GameOfThrones but this file was not created in the process.
3. Do we have to create a symbolic link to the database (audio .wav files) for ~/.pyannote/db.yml which redirects to /path/to/GameOfThrones/corpus/{uri}.wav
   • Is this {uri} variable need to be defined externally in bash terminal or is directly incorporated from files?

I went through the Speech Activity Detection tutorial succesfully, and the resulting test mdtm files look good in what it comes to basic segmentation, but there are some problems with Change Detection tutorial.

I get the same error now when I try evaluate and apply commands:

Using Theano backend.
Traceback (most recent call last):
  File "/Users/niko/anaconda3/envs/pyannote/bin/pyannote-change-detection", line 11, in <module>
    load_entry_point('pyannote.audio', 'console_scripts', 'pyannote-change-detection')()
  File "/Users/niko/pyannote_test/pyannote-audio/pyannote/audio/applications/change_detection.py", line 392, in main
    epoch=epoch, min_duration=min_duration)
  File "/Users/niko/pyannote_test/pyannote-audio/pyannote/audio/applications/change_detection.py", line 268, in evaluate
    predictions[uri] = aggregation.apply(dev_file)
  File "/Users/niko/pyannote_test/pyannote-audio/pyannote/audio/labeling/aggregation.py", line 120, in apply
    predictions = next(self.from_file(current_file))
  File "/Users/niko/anaconda3/envs/pyannote/lib/python3.5/site-packages/pyannote/generators/batch.py", line 365, in from_file
    incomplete=incomplete):
  File "/Users/niko/anaconda3/envs/pyannote/lib/python3.5/site-packages/pyannote/generators/batch.py", line 418, in __call__
    for fragment in self.generator.from_file(preprocessed_file):
  File "/Users/niko/anaconda3/envs/pyannote/lib/python3.5/site-packages/pyannote/generators/fragment.py", line 148, in from_file
    wav = current_file['wav']
KeyError: 'wav'

Thanks for help!

Hi, Bredin,

I am a newer to LSTM. After reading your paper(in the citation) about tristounet, I think I have got the basic idea how it works for speaker change detection.However, I still confused about the theory background for command: pyannote-speech-detection.

Can I say that speech detection is similar to speaker turns detection if non-speech segments are considered from a special 'speaker' while speech parts are from another 'speaker'. In this case, speech boundary detection is the same to speaker change detection.

However, I am still wondered why the parameter setting in config.yml ( n_classes: 2 ) for speech activity detection is different with the setting in config.yml ( n_classes: 1 ). To be honest, I don't know the meaning of this parameter(n_classes). Could I have other introduction or tutorials about the theory background for this pyannote-speech-detection command?

Thank you for your time and patience.

Liyong Guo

Hey,

Guess we both were looking forward to the Python 3 support on yaafe. Check it out, it works well for me:
https://github.com/Yaafe/Yaafe

We should implement (and evaluat) the following ones.
Given anchor and positive samples,

• hard negative - choose negative at random such that
  d(anchor, negative) < d(anchor, positive) + margin
• hardest negative - same as hard negative +
  negative = argmin(d(anchor, negative))
• semi-hard negative - same as hard negative +
  d(anchor, negative) > d(anchor, positive)

Hi,

We are trying to use pyannote for Speaker Change Detection in supervised way.

We have created our own database. We are trying to follow tutorial instructions. But getting following error. We haven't modified the code.

We have attached the screenshot of the error and our newly created database.

Could you please guide us in this?

Regards
Ankur
pyannote-db-template-master.tar.gz

Hi Hervé BREDIN,
what's the "architecture.yml" file of Pretrained ETAPE model?
"A TristouNet model (trained and validated on ETAPE database) is available for you to test directly in tutorials/speaker-embedding/2+0.5/TristouNet."

in TristouNet folder, I only find "0986.h5", maybe that's the weight_h5, but where is the architecture.yml?

• https://arxiv.org/pdf/1712.05015.pdf
• Exponential Discriminative Metric Embedding in Deep Learning
• A Discriminative Feature Learning Approach for Deep Face Recognition
• Ring loss: Convex Feature Normalization for Face Recognition
• Spoken Language Identification Using LSTM-Based Angular Proximity
• Focal Loss for Dense Object Detection
• Magnet Loss: Metric Learning with Adaptive Density Discrimination
• MTGAN: Speaker Verification through Multitasking Triplet
  Generative Adversarial Networks
• Sampling Matters in Deep Embedding Learning
• Angular Softmax Loss for End-to-end Speaker Verification
• Angular Softmax for Short-Duration Text-independent Speaker Verification
• Robust and Discriminative Speaker Embedding via Intra-Class Distance Variance Regularization
• NormFace: L2 Hypersphere Embedding for Face Verification
• Improved Large-margin Softmax Loss for Speaker Diarisation
• Powerful Speaker Embedding Training Framework by Adversarially Disentangled Identity Representation
• Divide and Conquer the Embedding Space for Metric Learning

Code

• https://github.com/cvqluu/Angular-Penalty-Softmax-Losses-Pytorch
  Angular penalty loss functions in Pytorch (ArcFace, SphereFace, Additive Margin, CosFace)

Hello, If I understand right, you don't use "metric" in module triplet_loss.py and it don't transfer SequenceEmbeddingAutograd after kwargs. In this way always use "cosine" as default.
Model traininig very slow due to distance computing. I have dataset with 1000+ speakers around 10 minutes for each and GPU. The first epoch's metrics were not computed within 3 hours. May be you have some idea how I can speed up it?
My approach settings is:

     metric: cosine
     margin: 0.1
     clamp: positive
     per_batch: 1
     per_label: 40
     per_fold: 40
     gradient_factor: 10000
     batch_size: 512

• https://github.com/idnavid/pyknograms
• https://github.com/idnavid/dissertation/blob/master/SHOKOUHI-DISSERTATION-2017-rev3.pdf
• LEVERAGING LSTM MODELS FOR OVERLAP DETECTION IN MULTI-PARTY MEETINGS

• Unsupervised Learning by Predicting Noise
• Unsupervised Learning in LSTM Recurrent Neural Networks
• Neural Predictive Coding using Convolutional Neural Networks towards Unsupervised Learning of Speaker Characteristics
• Unspeech: Unsupervised Speech Context Embeddings
• An overview of proxy-label approaches for semi-supervised learning
• Comparing the Max and Noisy-Or Pooling Functions in Multiple Instance Learning for Weakly Supervised Sequence Learning Tasks
• Learning Speaker Representations with Mutual Information
• Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations
• Recent Advances in Autoencoder-Based Representation Learning
• Self-supervised speaker embeddings
• Problem Agnostic Speech Encoder (PASE) YouTube video
• Representation Learning with Contrastive Predictive Coding
• Mockingjay: Unsupervised Speech Representation Learning with Deep Bidirectional Transformer Encoders

I have a one conversation with several people and it is unsupervised. I would only like to segment all the speakers in the conversation. In your speaker change detection code, i have given that that url(Local filepath) of the dataset. But after one interation I was getting an error which I believe due to the length of the dataset. Which part of the code I can use from this project.?

My final goal is to only segment unique speakers from the conversation. Any idea would be appreciated or any basic algorithmic steps which i can borrow from your code.

I guess this is just a matter of:

• setting return_sequences to True
• adding GlobalAveragePooling1D layers on top of LSTMs

I just want to experiment with the "Speaker Change Detection" as a SOLUTION OUT OF BOX ; for the first step, I need to test it on my own unlabeled data using a pretrained model and see if it generalizes well; and then decide whether to train my own model.

Something that looks like a pretrained model was find here:
https://github.com/yinruiqing/change_detection/tree/master/model

But I do not have access to the ETAPE database, nor do I understand the complicated structure or concepts (e.g. terms like "protocol" in "<database.task.protocol>") of the database;
So, how can I replicate your amazing work on my own data ?

Thanks !

Hi Bredin,

I want to cite your idea of "Binarize predictions using onset/offset thresholding" for speech activity detection. I think this is quite interesting comparing to a single threshold.

However, I failed to find introduction about that in two of your papers "Tristounet: Triplet Loss for Speaker Turn Embedding" and "Speaker Change Detection in Broadcast TV using Bidirectional Long Short-Term Memory Networks".

Could you tell me which paper of yours contains introduction about this idea so that I could cite in my work?

Many thanks.

Hello,
I try to run pyannote on my dataset, and I got error in pack_ndarray function of "pyannote/generators/batch.py"

def pack_ndarray(self, ndarrays):
    return np.stack(ndarrays)

I print shape of my ndarrays
(201, 35)
(201, 35)
(201, 35)
(201, 35)
(108, 35)
(201, 35)
(175, 35)
(157, 35)
(34, 35)
(51, 35)
...
the first dim is not the same, so what's my mistake?

• Speaker Diarization using Deep Recurrent Convolutional Neural Networks for Speaker Embeddings
• End-to-end text-independent speaker verification with flexibility in utterance duration
• VoxCeleb: a large-scale speaker identification dataset
• An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling
• Analysis of Length Normalization in End-to-End Speaker Verification System
  Authors show that normalizing embeddings on a larger sphere leads to better results.
• Speaker Recognition from Raw Waveform with SincNet
• Adaptive pooling operators for weakly labeled sound event detection
  This could be used to learn the final pooling layer in LSTM-based embedding
• TDNN

The original FaceNet paper uses squared euclidean distance in the triplet selection algorithm.

As of now, we use the regular euclidean distance:
See 9b59990#diff-25feed99c8806859b2dcd889d2dcc34fR137

It probably isn't big of a deal but changing euclidean to sqeuclidean would not only result in being identical to the original paper, but also speed up this part of the code.

Hi,
I have the latest version of Keras (2.1.3) and when I tried to train a model for speaker embedding I received the following error:

TypeError: get_updates() missing 1 required positional argument: 'constraints'.

I looked at keras' sources and this error is due to an API change in keras.optimizer.get_updates() that was not changed in the file pyannote/audio/optimizers.py. When I tried to downgrade to keras 2.0.0 I got another error related to inconsistency in keras.models.save_model signature.

Please fix this inconsistency or provide a suitable version of Keras.
Thanks!

This could be used like that

embedding.py train --validation=<subset> ...

this would result in a few files being created (such as evolution of EER as a function of epoch number, evolution of scores distribution, etc...)

cc @GregGovit

Hello! I have my own dataset with ~1500 speakers (3-10 minutes for each). I tried to train models with pyannote and have some results now.

1. Models works well on audio from the same devises (train and test stages), but the system quality falls very low on audio from different devices. I understand that it's may be actual only for my database. Do you test your model the same way or do you observe the same effects?
2. The quality grew when I have used 20 MFCC instead 11.
3. Pre-training model works on Russian language, EER around 20% (60 speakers).
4. I tried to fine-tune pre-training model. I used "--start" option for that, but got an error "model wasn't compiled". It fixed by adding:
   embedding.compile(optimizer=optimizer, loss=precomputed_gradient_loss)
   in 306 line base_autograd.py file, but I'm not sure that it's correct.

Thanks!

Hello Hervé,
I am currently trying to train the speaker embedding module using the TristouNet architecture but I end up with a loss of nan from the second epoch... So, here is the command I am running:

 $ pyannote-speaker-embedding-keras train --database=db.yml --subset=train tutorials/speaker-embedding/2+0.5/TristouNet Etape.SpeakerDiarization.TV

And here are the warnings/log messages:

/home/mahu/anaconda3/envs/pyannote/lib/python3.5/site-packages/h5py/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.
  from ._conv import register_converters as _register_converters
Using TensorFlow backend.
/home/mahu/anaconda3/envs/pyannote/lib/python3.5/site-packages/pyannote/generators/indices.py:84: UserWarning: 5 labels (out of 179) have less than 3 training samples.
  per_label=per_label))
Epoch 1/1000
2018-01-24 17:20:21.787683: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX
/home/mahu/anaconda3/envs/pyannote/lib/python3.5/site-packages/autograd/core.py:81: RuntimeWarning: divide by zero encountered in power
  result_value = self.fun(*argvals, **kwargs)
/home/mahu/anaconda3/envs/pyannote/lib/python3.5/site-packages/autograd/numpy/numpy_grads.py:84: RuntimeWarning: invalid value encountered in multiply
  anp.sqrt.defvjp(   lambda g, ans, vs, gvs, x : g * 0.5 * x**-0.5)
/home/mahu/anaconda3/envs/pyannote/lib/python3.5/site-packages/autograd/numpy/numpy_grads.py:46: RuntimeWarning: invalid value encountered in multiply
  unbroadcast(vs, gvs, g * y * x ** anp.where(y, y - 1, 1.)))
/home/mahu/anaconda3/envs/pyannote/lib/python3.5/site-packages/numpy/core/_methods.py:29: RuntimeWarning: invalid value encountered in reduce
  return umr_minimum(a, axis, None, out, keepdims)
/home/mahu/anaconda3/envs/pyannote/lib/python3.5/site-packages/numpy/core/_methods.py:26: RuntimeWarning: invalid value encountered in reduce
  return umr_maximum(a, axis, None, out, keepdims)
1/1 [==============================] - 36s - loss: 0.0535
Epoch 2/1000
1/1 [==============================] - 30s - loss: nan
Epoch 3/1000
1/1 [==============================] - 31s - loss: nan

Some minor details: as you may have guessed, I have slightly changed the options to the train method of pyannote-speaker-embedding-keras so that, like the data command, a path other than ~/.pyannote/db.yml could be specified for the db.yml file.
The various config.yml files (tutorial/speaker-embedding/config.yml and tutorial/speaker-embedding/2+0.5/TristouNet/config.yml) have the same content as what is given in the corresponding tutorial.

That said, another odd thing is that 2 progress indicators are printed when running

$ pyannote-speaker-embedding-keras data --database=db.yml --duration=2 --step=0.5 tutorials/speaker-embedding/ Etape.SpeakerDiarization.TV

as shown below

Training set: 0it [00:00, ?it/s]
Training set: 28it [02:57,  6.32s/it]
100%|████████████████████████████████████| 81433/81433 [00:37<00:00, 2148.18it/s]
Development set: 0it [00:00, ?it/s]
Development set: 9it [00:47,  5.32s/it]
100%|████████████████████████████████████| 23298/23298 [00:11<00:00, 2082.87it/s]
Test set: 0it [00:00, ?it/s]
Test set: 9it [00:50,  5.66s/it]
100%|████████████████████████████████████| 22815/22815 [00:10<00:00, 2132.08it/s]

So I don't really know if the problem I am facing comes from the training phase or from the data used to train the NN. Looking a bit around, the warnings given by autograd may be related to the bug reported here.
Have you encountered this problem before? And if so, how did you manage to circumvent it?

Cheers,
Mathieu

Edit:
After some printing, it appears that logs = self.loss_and_grad(batch, embedding) in pyannote/audio/embedding/approaches_keras/base.py, l. 333, yields a gradient with NaN values on the first epoch. However, I haven't been able to find the definition of loss_and_grad to narrow down the problem yet.

Hi, when I run pip install "pyannote.audio==0.3", I got the following error msg:

In file included from _pysndfile.cpp:471:0:
pysndfile.hh:55:21: fatal error: sndfile.h: No such file or directory
#include <sndfile.h>
^
compilation terminated.
error: command 'gcc' failed with exit status 1

Failed building wheel for pysndfile
Running setup.py clean for pysndfile
Failed to build pysndfile

Is there a trained TristouNet on ETAPE that I can use for domain adaption on my task of speaker turn detection?

https://github.com/fchollet/keras/blob/3dd3e8331677e68e7dec6ed4a1cbf16b7ef19f7f/keras/utils/training_utils.py#L56-L75

https://gist.github.com/fchollet/87e9a3e0539ce268222d1d597864c098

Can you please supply some documentation on speaker embedding and detection? It's not clear how to use this package. Moreover does it include diarization?

I get following error messages when trying to run the tune command on my adapted template. With training everything went as expected, I think. This creates tune.png and tune.yml files, but I assume this doesn't look as it should:

This is the error message:

Iteration No: 1 ended. Evaluation done at provided point.
Time taken: 3.3038
Function value obtained: 0.0000
Current minimum: 0.0000
Iteration No: 2 started. Evaluating function at random point.
/Users/niko/anaconda3/envs/py35-pyannote-audio/lib/python3.5/site-packages/keras/models.py:245: UserWarning: No training configuration found in save file: the model was *not* compiled. Compile it manually.
  warnings.warn('No training configuration found in save file: '
/Users/niko/anaconda3/envs/py35-pyannote-audio/lib/python3.5/site-packages/skopt/optimizer/optimizer.py:195: UserWarning: The objective has been evaluated at this point before.
  warnings.warn("The objective has been evaluated "
  ... (the same message repeated)
  Iteration No: 20 ended. Search finished for the next optimal point.
Time taken: 8.3426
Function value obtained: 0.0000
Current minimum: 0.0000
Exception ignored in: <bound method BaseSession.__del__ of <tensorflow.python.client.session.Session object at 0x12d522a58>>
Traceback (most recent call last):
  File "/Users/niko/anaconda3/envs/py35-pyannote-audio/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 582, in __del__
AttributeError: 'NoneType' object has no attribute 'TF_DeleteStatus'

In pyannote-audio/pyannote/audio/labeling/models.py.
The default value of bidirectional should be in
{False, 'ave', 'concat'} . Not True.

Is it possible to use this library for detecting and removing "hold music" from a given audio file, so that all that remains is the actual conversation between the people.

See http://ieeexplore.ieee.org/abstract/document/7953097/

line 33
databse or database ?

This is meant to keep track of notebooks that we may publicize in the tutorial section at some point.

• TristouNet-based speaker change detection

This could be used in the following way:

embedding.py train --restart=<epoch> ...

This also needs the following changes

• update SequenceEmbedding.to_disk method to also save the state of the optimizer
• update SequenceEmbedding.from_disk (or add a new method) to load a saved optimizer
• update SequenceEmbedding.fit to support preloaded optimizer

cc @GregGovit

may be it's because pandas
pyannote/parser/timeline/base.py, line 54-67
df = pandas.read_table(path,
delim_whitespace=True,
header=None, names=self.fields(),
comment=self.comment(),
converters=self.converters(),
keep_default_na=False, na_values=[],
)
could be fix if add dtype={'uri': object}

@GregGovit does your implementation of weighted ClopiNet (i.e. ClopiNet with a final fully-connected layer) contains a bias in the linear activation?

I'm finishing going through this tutorial, and everything has gone very well otherwise, but in the last extraction part I get following error.

It is very possible that I need to do some bigger changes to my setup or do some of the earlier tutorials again, as there seems to have been quite many changes. Anyway, maybe I'm missing something obvious here so I open an issue first. I updated pyannote-audio and others to the newest versions.

pyannote-speaker-embedding apply --step=2.0 $VALIDATION_DIR/development.eer.txt ikdp.SpeakerDiarization.MyFirstProtocol $OUTPUT_DIR

Using Theano backend.
/Users/niko/anaconda3/envs/pyannote/lib/python3.5/site-packages/keras/models.py:251: UserWarning: No training configuration found in save file: the model was *not* compiled. Compile it manually.
  warnings.warn('No training configuration found in save file: '
Development set: 0it [00:00, ?it/s]
Development set: 0it [00:00, ?it/s]Traceback (most recent call last):
  File "/Users/niko/anaconda3/envs/pyannote/bin/pyannote-speaker-embedding", line 11, in <module>
    load_entry_point('pyannote.audio', 'console_scripts', 'pyannote-speaker-embedding')()
  File "/Users/niko/pyannote_test/pyannote-audio/pyannote/audio/applications/speaker_embedding.py", line 791, in main
    internal=internal)
  File "/Users/niko/pyannote_test/pyannote-audio/pyannote/audio/applications/speaker_embedding.py", line 708, in apply
    fX = extraction.apply(current_file)
  File "/Users/niko/pyannote_test/pyannote-audio/pyannote/audio/embedding/extraction.py", line 142, in apply
    incomplete=True)])
  File "/Users/niko/pyannote_test/pyannote-audio/pyannote/audio/embedding/extraction.py", line 141, in <listcomp>
    [batch for batch in self.from_file(current_file,
  File "/Users/niko/anaconda3/envs/pyannote/lib/python3.5/site-packages/pyannote/generators/batch.py", line 379, in from_file
    incomplete=incomplete):
  File "/Users/niko/anaconda3/envs/pyannote/lib/python3.5/site-packages/pyannote/generators/batch.py", line 432, in __call__
    for fragment in self.generator.from_file(preprocessed_file):
  File "/Users/niko/anaconda3/envs/pyannote/lib/python3.5/site-packages/pyannote/generators/fragment.py", line 154, in from_file
    raise ValueError('source must be one of "annotated", "annotated_extent", "annotation", "support" or "audio"')
ValueError: source must be one of "annotated", "annotated_extent", "annotation", "support" or "audio"

Thanks for help!

what's the time cost during the training period?

Two advantages:

• control what is shown to the end user
• control which part of the code is run in parallel (needed to use GPU)

Hi,

I have two doubts :

1. Is 'speech activity detection' prerequisite to 'speaker change detection'?

2. Another question is current approach in pyannote is speaker and content invariant?

3. In striking balance between purity and coverage, what should be the value of coverage that should be good enough practically?

Regards
Ankur

This looks really nice. I was wondering if you have or plan to make available a quickstart example using an open source database?

There is a breaking change in Keras optimizer in version 2.0.7. So when I use a SMORMS3 (an optimizer defined in pyannote-audio and used in pyannote-speech-detection and pyannote-change-detection), it raise an error:
get_updates() missing 1 required positional argument: 'constraints'

ValidationCheckpoint should save the best model so far in something like best.accuracy.h5 and best.fscore.h5

It should also remember which epoch it corresponds to so it can be plotted in status.png.