This guide is written to help you get the tool working on your machine. We will use a example setup that involves training a phoneme transcription tool for Yongning Na. For this we use a small (even by language documentation standards) sub-sampling of elicited speech of Yongning Na, a language of Southwestern China.

The example that we will run can be run on most personal computers without a graphics processing unit (GPU), since I’ve made the settings less computationally demanding than it would be for optimal transcription quality. Ideally you’d have access to a server with more memory and a GPU, but this isn’t necessary.

The code has been tested on Mac and Linux systems. It can be run on Windows using the Docker container described below.

For now you must open up a terminal to enter commands at the command line. (The commands below are prefixed with a ":math:`". Don’t enter the "`", just whatever comes afterwards).

1. Installation

Installation option 1: Using the Docker container

To simplify setup and system dependencies, a Docker container has been created. This just requires Docker to be installed. Once you have installed docker you can fetch our container with:

$ docker pull oadams/persephone

Then run it in interactive mode:

$ docker run -it oadams/persephone

This will place you in an environment where Persephone and its dependencies have been installed, along with the example Na data.

Installation option 2: A “native” install

Ensure Python 3 is installed.

You will also need to install some system dependencies. For your convenience we have an install script for dependencies for Ubuntu. To install the Ubuntu binaries, run ./ubuntu_bootstrap.sh to install ffmpeg packages. On MacOS we suggest installing via Homebrew with brew install ffmpeg.

We now need to set up a virtual environment and install the library.

$ python3 -m virtualenv -p python3 persephone-venv
$ source persephone-venv/bin/activate
$ pip install -U pip
$ pip install persephone
$ pip install ipython

(This library can be installed system-wide but it is recommended to install in a virtualenv.)

I’ve uploaded an example dataset that includes some Yongning Na data that has already been preprocessed. We’ll use this example dataset in this tutorial. Once we confirm that the software itself is working on your computer, we can discuss preprocessing of your own data.

Create a working directory for storage of the data and running experiments:

mkdir persephone-tutorial/
cd persephone-tutorial/
mkdir data

Get the data here

Unzip na_example_small.zip. There should now be a directory na_example/, with subdirectories wav/ and label/. You can put na_example anywhere, but for the rest of this tutorial I assume it is in the working directory: persephone-tutorial/data/na_example/.

2. Training a toy Na model

One way to conduct experiments is to run the code from the iPython interpreter. Back to the terminal:

$ ipython
> from persephone import corpus
> corp = corpus.Corpus("fbank", "phonemes", "data/na_example")
> from persephone import experiment
> experiment.train_ready(corp)

You’ll should now see something like:

Number of training utterances: 1024
Batch size: 16
Batches per epoch: 64
2018-01-18 10:30:22.290964: 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 AVX2 FMA
exp_dir ./exp/0, epoch 0

The message may vary a bit depending on your CPU but if it says something like this then training is very likely working. Contact me if you have any trouble getting to this point, or if you had to deviate from the above instructions to get to this point.

On the current settings it will train through at least 10 “epochs”, very likely more. If you don’t have a GPU then this will take quite a while, though you should notice it converging in performance within a couple hours on most personal computers.

After a few epochs you can see how its going by going to opening up exp/<experiment_number>/train_log.txt. This will show you the error rates on the training set and the held-out validation set. In the exp/<experiment_number>/decoded subdirectory, you’ll see the validation set reference in refs and the model hypotheses for each epoch in epoch<epoch_num>_hyps.

Currently the tool assumes each utterance is in its own audio file, and that for each utterance in the training set there is a corresponding transcription file with phonemes (or perhaps characters) delimited by spaces.

3. Using your own data

If you have gotten this far, congratulations! You’re now ready to start using your own data. The example setup we created with the Na data illustrates a couple key points, including how your data should be formatted, and how you make the system read that data. In fact, if you format your data in the same way, you can create your own Persephone Corpus object with:

corp = corpus.Corpus("fbank", "phonemes", "<your-corpus-directory>")

where extension is “txt”, “phonemes”, “tones”, or whatever your file has after the dot.

If you are using the Docker container then to get data in and out of the container you need to create a “volume” that shares data between your computer (the host) and the container. If your data is stored in /home/username/mydata on your machine and in the container you want to store it in /persephone/mydata then run:

docker run -it -v /home/username/mydata:/persephone/mydata oadams/persephone

This is simply an extension of the earlier command to run docker, which additionally specifies the portal with which data is transferred to and from the container. If Persephone—abducted by Hades—is the queen of the underworld, then you might consider this volume to be the gates of hell.

Formatting your data

Interfacing with data is a key bottleneck in useability for speech recognition systems. Providing a simple and flexible interface to your data is currently the most important priority for Persephone at the moment. This is a work in progress.

Current data formatting requirements:

  • Audio files are stored in <your-corpus>/wav/. The WAV format is supported. Persephone will automatically convert wavs to be 16bit mono 16000Hz.
  • Transcriptions are stored in text files in <your-corpus>/label/
  • Each audio file is short (ideally no longer than 10 seconds). There is a script added by Ben Foley, persephone/scripts/split_eafs.py, to split audio files into utterance-length units based on ELAN input files.
  • Each audio file in wav/ has a corresponding transcription file in label/ with the same prefix (the bit of the filename before the extension). For example, if there is wav/utterance_one.wav then there should be label/utterance_one.<extension>. <extension> can be whatever you want, but it should describe how the labelling is done. For example, if it is phonemic then wav/utterance_one.phonemes is a meaningful filename.
  • Each transcript file includes a space-delimited list of labels to the model should learn to transcribe. For example:
    • data/na_example/label/crdo-NRU_F4_ACCOMP_PFV.0.phonemes contains l e dz ɯ z e l e dz ɯ z e
    • data/na_example/label/crdo-NRU_F4_ACCOMP_PFV.0.phonemes_and_tones might contain: l e ˧ dz ɯ ˥ z e ˩ | l e ˧ dz ɯ ˥ z e ˩
  • Persephone is agnostic to what your chosen labels are. It simply tries to figure out how to map speech to that labelling. These labels can be multiple characters long: the spaces demarcate labels. Labels can be any unicode character(s).
  • Spaces are used to delimit the units that the tool predicts. Typically these units are phonemes or tones, however they could also just be orthographic characters (though performance is likely to be a bit lower: consider trying to transcribe “$100”). The model can’t tell the difference between digraphs and unigraphs as long as they’re tokenized in this format, demarcated with spaces.

If your data observes this format then you can load it via the Corpus class. If your data does not observe this format, you have two options:

  1. Do your own separate preprocessing to get the data in this format. If you’re not a programmer this is probably the best option for you. If you have ELAN files, this probably means using persephone/scripts/split_eaf.py.
  2. Create a Python class that inherits from persephone.corpus.Corpus and does all your preprocessing. The API (and thus documentation) for this is work in progress, but the key point is that <corpusobject>.train_prefixes, <corpusobject>.valid_prefixes, and <corpusobject>.test_prefixes are lists of prefixes for the relevant subset of the data. For an example on a full dataset, see at persephone/datasets/na.py (beware: here be dragons).

Creating validation and test sets

Currently Corpus splits the supplied data into three sets (training, validation and test) in a 95:5:5 ratio. The training set is what your model is exposed to during training. Validation is a held-out set that is used to gauge during training how well the model is performing. Testing is what is used to quantitatively assess model performance after training is complete.

When you first load your corpus, Corpus randomly allocates files to each of these subsets. If you’d like to do change the prefixes of which utterances are in in each set, modify <your-corpus>/valid_prefixes.txt and <your-corpus>/test_prefixes.txt. The training set consists of all the available utterances in neither of these text files.

4. Miscellaneous Considerations

On choosing an appropriate label granularity


Suprasegmentals like tone, glottalization, nasalization, and length are all phonemic in the language I am using. Do they belong in one grouping or separately?


I’m wary of making sweeping claims about the best approach to handle all these sorts of phenomena that will realise themselves differently between languages, since I’m neither a linguist nor do I have strong understanding for what features the model will learn each situation. (Regarding tones, the literature on this is also inconclusive in general). The best thing is to empirically test both approaches:

  1. Having features as part of the phoneme token. For example, a nasalized /o/ becomes /õ/.
  2. Having a separate token that follows the phoneme. For example, a high tone /o˥/ becomes two tokens: /o ˥/.

Since there are many ways you can mix and match these, one consideration to keep in mind is how much larger the label vocabulary becomes by merging two tokens into one. You don’t want this vocabulary to become too big because then its harder to learn features common to different tokens, and the model is less likely to pick the right one even if it’s on the right track. In the case of vowel nasalization, maybe you only double the number of vowels, so it might be worth having merged tokens for that. If there are 5 different tones though, you might make that vowel vocabulary about 5 times bigger by combining them into one token, so its less likely to be good (though who knows, it might still yield performance improvements).

5. Saving and loading models; transcribing untranscribed data

So far, the tutorial described how to load a Corpus object, and perform training and testing with a single function run.train_ready(corpus), which hid some details. This section exposes more of the interface so that you can describe models more fully, save and load models, and apply it to untranscribed data. I’d like to hear people’s thoughts on this interface.

CorpusReaders and Models

The Corpus object, is an object that exposes the files in the corpus (among several other things). Of relevance here is the .get_train_fns(), .get_valid_fns(), .get_test_fns() methods, which provide lists of files in the training, validation and test sets respectively. There is additionally a .get_untranscribed_fns() method which returns a list of files representing speech that has not been transcribed. .get_untranscribed_fns() fetches prefixes of utterances from untranscribed_prefixes.txt, which you can put in the corpus data directory (at the same level as the feat/ and label/ subdirectories).

To fetch data from your Corpus, a CorpusReader is used. The CorpusReader regulates how much data is to be read from the corpus, as well as the size of the “batches” that are fed to the model during training. You create a CorpusReader by feeding it a corpus (here the example na_corpus):

from persephone import corpus
na_corpus = corpus.Corpus("fbank", "phonemes", "data/na_example/")
from persephone import corpus_reader
na_reader = corpus_reader.CorpusReader(na_corpus, num_train=512, batch_size=16)

Here, na_reader is an interface to the corpus which will read from the corpus files 512 training utterances, in batches of 16 utterances. We can now feed data to a Model:

from persephone import rnn_ctc
model = rnn_ctc.Model(exp_dir, na_reader, num_layers=2, hidden_size=250)

where exp_dir is a directory in which experimental results and logging will be stored. In creating an rnn_ctc.Model (recurrent neural network with a connectionist temporal classification loss function) we have also specified what corpus to read from, how many layers there are in the neural network, and the amount of “neurons” in those layers. We can now train the model with:


After training, we can transcribe untranscribed data with:


which depends on untranscribed_prefixes.txt existing before corpus creation (though there’s no reason why this can’t be changed to simply transcribe the utterances with feature files in <data-dir>/feat/ that don’t have corresponding transcriptions in <data-dir>/label/).

During training, the model will store the model that performs best on the validation set in <exp_dir>/model, across a few different files prefixed with model_best.ckpt. If you later want to load this model to transcribe untranscribed data, you create a model with the same hyperparameters and call model.transcribe() with the restore_model_path keyword argument:

model = rnn_ctc.Model(<new-exp-dir>, na_reader, num_layers=2, hidden_size=250)

This will load a previous model and perform transcription with it.