About the MATERIAL corpus:

The MATERIAL project:
https://www.iarpa.gov/index.php/research-programs/material
https://www.nist.gov/itl/iad/mig/openclir-evaluation

The speech data in the MATERIAL corpus consist of four data sets for each
language: train (BUILD), development (BUILD-dev), test (ANALYSIS1 and ANALYSIS2),
and unlabeled evaluation audio (EVAL{1,2,3}). The train, development, test, and
evaluation data contain around 40, 10, 20, and 250 hours of audio respectively.
The train set is transcribed conversational audio that can be used for training
an ASR system. It consists of some in 8-bit a-law .sph (Sphere) files and some
in .wav files with 24-bit samples. The development set is transcribed
conversational audio that can be used as development data for training to tune
model parameters. The test data come in long unsegmented files. The reference
transcripts for the test set is provided, hence, one can measure WER on the test
set. The evaluation set is untranscribed audio that can be used for
semi-supervised training of the acoustic model.
Conversational speech data in the train and test sets are two-channel audio with
the two channels temporally aligned. Each audio channel is provided and
transcribed as a separate file, identified as inLine or outLine channel. Both
audio channels are interleaved in a single file and a there is a single
interleaved transcript that reflects the temporal alignments. In addition to
conversational speech, the test and evlatuion sets also contain other
genres of speech, namely news broadcast and topical broadcast, which are
single channel files.


Running the recipe:

In s5)
./run.sh --language <swahili|tagalog|somali>
./local/chain/run_tdnn.sh
./local/chain/decode_test.sh --language <swahili|tagalog|somali>
./local/rnnlm/run_tdnn_lstm.sh