A set of tools to generate training data suitable for text-to-speech systems out ot the recorded parliamentary meetings and their stenographic transcripts.

This dataset is quite large (about 200GB downloadable mp3's) and in public domain. The transcripts are manual but not 100% literal, though the precision is very good. The major downsides are a complete lack of machine-readable metadata and only very loose alignment.

There is a small set of these transcribed recordings made available as a richly annotated corpus for speech recognition by the University of West Bohemia in Pilsen. The set is available through Lindat / Clarin under the name "Czech Parliament Meetings" and it can be used to train a seed model.

The goal is to have couples wav-file -- txt-file under 30 seconds of length. What we start with is a website with HTML-formatted text and links to mp3's. The scheme is as follows:

1. scrape the transcripts coupled to their corresponding mp3 urls,
2. download all the mp3's,
3. train a seed model using the Pilsen dataset,
4. automatically transcribe the mp3's using the seed model keeping word-level forced alignment,
5. align the automatic transcript with the manual one,
6. generate chunks aligned with reliable manual transcripts.

Install scrapy to scrape the transcripts and URLs of their corresponding audio recordings.

virtualenv -p python3 parliament
. parliament/bin/activate
pip install scrapy python-levenshtein

Install julius for the initial aligned transcript.

You'll need the HTK. (HMM ToolKit) if you want to train your own acoustic model for the initial transcription. You'll need the program HCopy from HTK to run the initial transcription.

To get the initial transcript, you need julius and an acoustic and a language model. You can either download the model package and extract it here or you can build it yourself with HTK.

Please install a recent enough version of wget, so it can cope with HTTPS redirects.

Install sox with mp3 support.

Install kenlm so that the binary lmplz is in the PATH.

Get the Perl module JSON::XS.

The recordings and their transcripts are available from the website of the parliament: psp.cz/eknih.

The transcripts and their corresponding mp3 URLs can be scraped from the parliament web by the scrapy scraper.

scrapy runspider -o downloads/scraped.json scripts/scrape.py

To download the mp3's, use download-scraped-audio.sh:

bash scripts/download-scraped-audio.sh downloads/scraped.json

Never mind the 404s. The earlier recordings simply seem to be missing. The result should be about 17000 files. The download will likely take a few hours or even days depending on your connection quality.

The mp3s will be downloaded into the directory downloads/audio. If you want to download it someplace else, please use a symlink.

You'll need a trigram language model in ARPA format if you plan to

1. make an initial aligned transcription,
2. build an acoustic model for the purposes of initial transcription,
3. use the final model with an engine that requires one, like Mozilla's DeepSpeech.

So you probably do need a language model. You can download it along with the acoustic model or you can build it yourself.

To build the language model, you'll need the scraped data from the above section (downloads/scraped.json) and the package kenlm.

bash scripts/mklm.sh

An initial transcription is necessary for alignment of the stenographs with the recordings. We use HTK for training, Julius for inference and the Parliament Meetings corpus released by University of West Bohemia through Lindat/Clarin as training data.

If you prefer to build your own acoustic model for the initial transcription, see the directory htkasr.

Once you have the acoustic and language model (trained or downloaded), you can proceed to use julius to get the initial transcript:

. config.sh
bash scripts/initial-transcription.sh

Note that this stage takes by far the longest. It will use all the processors available. You can affect the number of processors used by setting the environment variable PROCESSOR_COUNT.

Generating couples of sentence-level audio files and their transcripts is the final step in preparing engine-agnostic data for speech recognition. So far we have original 14-minute MP3's, manual transcripts corresponding roughly to middle 10 minutes, and word-level-aligned automatic transcripts. Hence, we 1. create an alignment between the audio-aligned automatic transcript and the manual transcript, 2. split to chunks surrounded by silences, 3 select reliable chunks and 4. generate their wav-text file couples.

It is first necessary though to extract the transcripts:

export OUTDIR=data/extracted
perl scripts/extract-transcript.pl < downloads/scraped.json

Then run ls data/recout/* | bash scripts/split.sh.

The results are stored in data/splits. Make sure there is enough space in this directory, as the corpus will be saved in 16kHz wav format (2:1 to original mp3 files).

For more detail on aligning, see README-aligning.md.