This document contains instructions to download and use ASR models tailored for child speech. It is well known that child speech is harder for computers to understand when compared to adult speech. While a number of contributing factors have been identified and worked upon, child WERs are usually significantly higher than adult WERs. Further, freely available child speech corpora are few in number and scarce.
At the moment, there are two DNN-HMM hybrid models available based on time-delay neural network (TDNN) and self-attention. Both models were trained using 2700 hours (clean speech: 700 hrs, noise and reverb augmented: 2000 hrs) of prompted and spontaneous child speech. The models are benchmarked using a subset of the CIDMIC corpus, containing 3.2 hours of prompted speech from children between the ages 5 and 17 years
| Model | Deepspeech | ASPIRE | TDNN | Attention |
|---|---|---|---|---|
| WER | 52.32 | 36.38 | 29.50 | 29.76 |
Download the archive from http://bit.ly/2PKkuP3 and extract it. You will find the following:
child_asr/ cmd.sh decode.sh libri_data/ libri_demo/ mfcc.conf path.sh
The following variables need to be set in decode.sh:
- suffix : A string, to name your test data
- dataDir : Kaldi data directory
- kaldiPath: Path to Kaldi installation
You might need to install flac to decode the demo data.
Manoj Kumar (prabakar@usc.edu)