End-to-end spoken language identification (LID) on TensorFlow
- End-to-end spoken language identification (LID) on TensorFlow.
- Parallel feature extraction using
tf.data.Dataset, with STFT computations on the GPU using the
- Only metadata (e.g. utt2path, utt2label) is fully loaded into memory, rest is done in linear passes over the dataset with the
- Spectrograms, source audio, and utterance ids can be written into TensorBoard summaries.
- Model training with
tf.keras, some model examples are available here.
- Average detection cost (
C_avg) implemented as a
- You can also try
lidboxfor speaker recognition, since no assumptions will be made of the signal labels. E.g. use utt2speaker as utt2label and see what happens.
Here is a full example notebook showing what
lidbox can do.
Why would I want to use this?
- You need a simple, deep learning based speech classification pipeline. For example: waveform -> VAD filter -> augment audio data -> serialize all data to a single binary file -> extract log-scale Mel-spectra or MFCC -> use DNN/CNN/LSTM/GRU/attention (etc.) to classify by signal labels
- You have thousands of hours of speech data
- You have a TensorFlow/Keras model that you train on the GPU and want the
tf.data.Datasetextraction pipeline to also be on the GPU
- You want an end-to-end pipeline that uses TensorFlow 2 as much as possible
Why would I not want to use this?
- You are happy doing everything with Kaldi or some other toolkits
- You don't want to debug by reading the source code when something goes wrong
- You don't want to install TensorFlow 2 and configure its dependencies (CUDA etc.)
- You need CTC or some other way to train a phoneme recognizer
git clone --depth 1 https://github.com/matiaslindgren/lidbox.git pip install ./lidbox
Check that the command line entry point is working:
If not, make sure the
setuptools entry point scripts (e.g. directory
$HOME/.local/bin) are on your path.
Then, install TensorFlow 2.1 or 2.2 (both should work), unless it is already installed.
If everything is working, see this for a simple example to get started.
If you want to use language embeddings, install the PLDA package from here:
pip install plda@https://github.com/matiaslindgren/plda/archive/as-setuptools-package.zip#egg=plda-0.1.0
If you plan on making changes to the code, it is easier to install
lidbox as a Python package in setuptools develop mode:
git clone --depth 1 https://github.com/matiaslindgren/lidbox.git pip install --editable ./lidbox
Then, if you make changes to the code, there's no need to reinstall the package since the changes are reflected immediately.
Just be careful not to make changes when
lidbox is running, because TensorFlow will use its
autograph package to convert some of the Python functions to TF graphs, which might fail if the code changes suddenly.
One benefit of deep learning classifiers is that you can first train them on large amounts of data and then use them as feature extractors to produce low-dimensional, fixed-length language vectors from speech. See e.g. the x-vector approach by Snyder et al.
Below is a visualization of test set language embeddings for 4 languages in 2-dimensional space. Each data point represents 2 seconds of speech in one of the 4 languages.
Release history Release notifications | RSS feed
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.