DynaSpeak® speech recognition engine


Features and benefits

  • Hidden Markov Model (HMM)-based speech recognizer – State of the art accuracy
  • Continuous speech – No need for pauses, user speaks naturally
  • Dynamic grammar compilation – Enables complex application workflows in a small
    footprint speech recognizer
  • Speaker independent – No tedious user training session required
  • Speaker adaptation – Automatically adjusts to different speakers and accents
  • C++ implementation – Portable to a range of hardware/software configurations
  • Noise filtering design tools – Rapid tuning for noisy acoustic environments without time-
    consuming and expensive acoustic model development
  • Dynamic noise compensation – Realtime differentiation between background noise and
  • Floating point or integer versions – Wide choice of hardware options
  • Supports finite state (command and control) or statistical (free form) grammars – More flexible, natural application designs
  • Supports push-to-talk, hold-to-talk, and open mic recording – Multiple user interface
  • Distributed speech recognition over low-bandwidth networks – Low-cost, high-accuracy
    deployment option for speech recognition on mobile devices

Technical specifications

CPU requirements

200 MHz StrongArm, 66 MHz Intel x86 (support for other processors on request)

Memory requirements

  • Total: 750KB-2.250MB
  • Executable (ROM): 350-750KB
  • Acoustic models (ROM): 100-500KB
  • Active search (RAM): 300KB-1MB (more for complex grammars)

Supported languages

  • Adults: American and British English, Latin American Spanish, Iraqi Arabic, Pashto and
    Dari (others on request)
  • Children: American English


  • Statistical or JSGF forms; static or dynamic; dictation option

Operating systems

  • Windows, Mac OS X, Linux and Android (others on request)

Development environment

  • C/C++, Java (via JNI), client/server versions available

Read more from SRI