Pop Up Archive Kaldi release

Prequisites

• [Kaldi] (https://github.com/kaldi-asr/kaldi)
• [SoX] (http://sox.sourceforge.net/) for audio files
• [ffmpeg] (https://www.ffmpeg.org/) for converting video files
• CMUSeg (use sample_experiment/install-cmuseg.sh)
• IRSTLM (use [KALDI-PATH]/tools/extras/install_irstlm.sh)
• sclite (see [KALDI-PATH]/tools/sctk-2.4.10)
• Latest version of [exp dir] (https://sourceforge.net/projects/popuparchive-kaldi/files/), should be (re-)namedexp

##Notes about Kaldi It is recommended that you review the [Kaldi documentation] (http://kaldi-asr.org/doc/) before you begin, especially if you intend to modify the compiled model included on Sourceforge.

Expected Set-up

Each model you experiment with should have its own directory. Start by putting the exp dir from Sourceforge in the sample_experiment dir.

Preliminaries for your experiment dir (e.g. sample_experiment)

• Make sure path.sh and set-kaldi-path.sh match your Kaldi location.
• Create the following sym links in your experiment dir
  • ln -s [KALDI-PATH]/egs/wsj/s5/steps [EXPT-DIR]
  • ln -s [KALDI-PATH]/egs/wsj/s5/utils [EXPT-DIR]

Other preliminaries

• Create a directory to store results from sclite, e.g. mkdir results
• Convert audio to mono 16-bit Signed Integer PCM, sample rate 16K
• Example with sox sox input.mp3 -c 1 -r 16000 -L output.wav
• Feel free to increase nj and/or decode_nj in exp/run.sh depending on how much memory you have. In decoding, expect each job to use at least 5 gb.
• If you have ground truth transcripts and wish to evaluate the accuracy of your output, create a text file in the following format:
  • The transcript for each file should be on a single line, all lower case without punctuation.
  • The line of transcript text should end with ([FILENAME]-1). Do not use dashes or underscores in the file name.
  • Include a new line at the end of the file.

Usage

Run kaldi speech recognition on directory of wav files: python run_kaldi.py [EXPERIMENT-DIR] [WAV-DIR]
Run evaluation ([RESULTS-DIR] will be created and sclite files will be written there): python run_sclite.py [KALDI-OUTPUT-DIR] [RESULTS-DIR] [REF-FILE-PATH]

##Building on the model You can use the current model as is, or add your own lexicon or language model.

• You can add new words to the lexicon by editing exp/dict/lexicon.txt and running sh prep_lang_local.sh exp/dict exp/tmp_lang exp/lang. Make sure your pronunciations only use phones that are already in the lexicon.
• Use sh add_grammar.sh [LM-FILEPATH] to create a new bigram language model based on an LM textfile and to update the overall model. Use sh create_big_lm.sh [LM-FILEPATH] lang_newlm lang_lmrescore to create a new 5-gram language for rescoring.
• Additional raw LM text can be obtained from [Open SLR] (http://www.openslr.org/27/). They also include trigram and 4-gram pre-trained LMs, however we recommend using a bigram baseline and 5-gram rescoring.

Acknowledgements

This repo is based on [work] (https://github.com/APMG/audio-search) done by APM with Cantab Research.