What's new in SPRAAK V1.2.38X

Compatibility with previous versions of SPRAAK

• UTF-8 is now used as internal character encoding (messages, ...) instead of latin1
• modified beam pruning; this may affect the results when using an aggressive pruning
• using less precision when evaluation log(), exp() where we assume the input is not all that precise (e.g. acoustic model evaluation because the input samples are 16 bit only)

New features

acoustic modelling

• Fast CPU-based implementation for DNN-based acoustic models; no training or GPU support (yet)
• Support for acoustic scores in various formats: linear or log or log10 or log<base>, posteriors or (scaled)likelihoods, normalized or not, ...
• Very fast routines to convert the acoustic scores from one of the formats to another (e.g. fast normalization of the acoustic scores).

language modelling

• Module for on-the-fly LM combination.
• The FSG LM now also supports LM extensions.

lattice processing

Routine to compute approximate confidence scores (combining the normalized acoustic score, the forward LM score, the backward LM score, ...).

improved pruning

• Options to reduce search effort on sequences of frames with high acoustic entropy

  See Also

  the "search adapt" command in spr_cwr_main.c

• Somewhat better beam width control (stricter pruning around the target number of tokens).

other changes in fuctionality

• Support for Windows
• Support for environment variable substitutions (${var}, and more complex constructions) in preproc and other ini/config files –

  See Also

  spr_feval_var_substitute() for details

• Support for math expression substitutions, i.e. $(<expr>) in preproc and other ini/config files –

  See Also

  spr_feval_var_substitute() for details

• Support for word/syllable internal triphones.
• Support for processor specific optimizations at compile time (native,

  See Also