|
Most speech application developers can "tune" an application to some extent, for instance by adding keywords to a grammar or fixing a phonetic pronunciation. At the same time, most don't realize that effective tuning goes so much beyond that. The reality is that effective tuning is a science that can only be done with deep expertise, sophisticated tools , and a rigorous data-driven methodology.
An effective tuning project needs:
- A rigorous experimental setup – Without benchmark tests that accurately model application performance in the field and optimization metrics that closely correlate with field performance, most of the tuning effort can turn out to be useless or may even end up deteriorating application performance.
- The ability to rapidly identify issues to be addressed – Because it's always necessary to limit the amount of time spent tuning an application, it's critical that we can rapidly identify the problems having the most significant impact on application performance and usability. Without sophisticated analysis tools, significant problems can easily go undetected when dealing with large test corpora.
- A sophisticated arsenal of tuning techniques and technologies – There's much more to tuning than changing grammar coverage and tweaking pronunciations. Effective tuning needs a large toolbox of techniques, including:
- Grammar analysis tools that can not only compute usage statistics for all alternatives, but also indicate those that contribute most to recognition errors.
- Out-of-grammar modeling techniques that minimize false accepts.
- Grammar development approaches that reduce pruning errors.
- The use of results-specific thresholds.
- The optimization of confidence thresholds through dialog simulation.
- The computation of enhanced confidence scores that are significantly superior those produced by the recognition engine
- The ability to analyze the impact of a change – Tuning is done through an iterative experimental process where the impacts of the latest changes need to be carefully analyzed and understood in order to make sure that the process will converge. This requires tools that provide powerful comparison analysis capabilities.
- Systematic regression testing – This ensures that no problem that can be detected through experimentation and simulations finds its way into production.
Nu Echo's tuning factory is absolutely unique in the industry. The difference speaks for itself. It is, of course, possible that your application may not require the best possible performance. But if it does, you should seriously consider contacting us.
|