kaldi-alligner

kaldi-alligner copied to clipboard

Kaldi Aligner: A simple script to create time alignment for given speech/transcription pairs. This script also enrich the transcription using [laughter] and [noise] markers. It does not use the forced-alignment instead it creates a bigram LM using the input transcription (after enriching the transcription with markers). After creating a language model it create an HCLG graph and use Kaldi decoder to generate a lattice and finally use the lattice to obtain alignment information.

requirements: 1- Kaldi tool. From: https://github.com/kaldi-asr/kaldi 2- SRILM (also existed under Kaldi/tools). From: http://www.speech.sri.com/projects/srilm/ 3- Python 3 4- bash (only tested under Linux)

After installing Kaldi and SRILM: open path.sh and update export KALDI_ROOT=/home/amir/Projects/kaldi to your kaldi path. Also make sure SRILM binaries (specifically ngram-count) is in the PATH.

Before running the aligner: Download the pretrained Aspire chain model by running: sh sownload_extract.sh This script downloads the model and also run some commands for preparation. Alternatively, you can train your own model. However, you might need to update the scripts accordingly.

Example: bash align.sh example/trans.txt example/test.wav data/lang_chain/ out.ctm out_phone.ctm out_transid_seq.txt lpf.txt

cat out.ctm test.wav 1 0.070 0.840 [noise] test.wav 1 0.910 0.320 my test.wav 1 1.240 0.300 name test.wav 1 1.540 0.340 is test.wav 1 1.880 0.300 [noise] test.wav 1 2.180 0.780 [laughter] test.wav 1 2.960 0.600 [noise] test.wav 1 3.630 0.360 amir test.wav 1 4.000 0.480 test.wav 1 4.510 1.610 [noise]

Notice OOVs are replaced with . The scripts adds [noise]/[laughter] markers when needed.