A Cache-Based Natural Language Model for Speech Recognition
IEEE Transactions on Pattern Analysis and Machine Intelligence
Digital Investigation: The International Journal of Digital Forensics & Incident Response
| Hi-index | 0.00 |
We present a method by which to determine a synchronzation point within a DEFLATE-compressed bit stream (as used in Zip and gzip archives) for which the beginning is unknown or damaged. Decompressing from the synchronization point forward yields a mixed stream of literal bytes and co-indexed unknown bytes. Language modeling in the form of byte trigrams and word unigrams is then applied to the resulting stream to infer probable replacements for each co-indexed unknown byte. Unique inferences can be made for approximately 30% of the co-indices, permitting reconstruction of approximately 75% of the unknown bytes recovered from the compressed data with accuracy in excess of 90%. The program implementing these techniques is available as open-source software.