Texas Instruments

Speech Recognizer Project  0_90_00_00

Release Notes 

Document Revision: 062416.1152

Introduction

The Speech Recognizer Project provides a low-resource speech recognition library called MinHMM.
The details of the MinHMM Library are provided in the documentation.

This document is divided into the following sections:

Documentation

What's New

0_90_00_00

Revision_History

0_80_00_00

Installation and Usage

Stand alone installation

  1. Download the SpeechRecognizer software
  2. Run the appropriate platform specific installer: SpeechRecognizer-<version>-<platform>-installer
  3. Follow the installer directions
  4. In the installation directory

Device Support

The following families of devices are supported in this release of Speech Recognizer Project :
The MinHMM library and MinHMMDemo program have been built for the toolchains in the following chart. When toolchain debugging support is available, the  library has been tested with the specified debugger. Although tested with a specific debugger,  code examples and the MinHMM library will work on a variety of different debuggers.

ToolchainDebugger
Texas Instruments Code Composer Studio 6.1.2TI XDS110
IAR Embedded Workbench for ARM 7.60.2TI XDS110 Emulator

Resources and Specifications

This section provides estimates of the resources needed for operation of MinHMM.

Memory Resources
The MinHMM library code size is about 23kB. The memory required to store model data depends on the maximum model size required. This release uses one flash sector of 4kB per model. MinHMM requires approximately 3kB of RAM that must persist as long as MinHMM is being used. During model training and recognition search additional temporary RAM is required for computation. For creating (enrolling) a model 4kB temporary RAM is required. For updating a model 20kB of temporary RAM is needed. For recogintion search 250 bytes per model is required.
 
Processing Cycles
Processor cycle requirements are a complex function of the trained models, the input speech data and background noise, and the quality of the match between the trained data and the input speech. As a general rule, when there is less background noise and the input speech matches a model well, then cycle usage decreases during active recognition search. When there is no speech and background noise is low or at a moderate level and not changing appreciably, MinHMM will assume the signal is in a background noise condition and enter a background processing mode. When speech is present, or when other background noises are appreciable and varying, then MinHMM will enter the active recognition search mode to attempt to match the audio signal to a model, thus causing higher cycle usage. The table below provides estimates of  typical cycle usage measured while running the MinHMM example demo program on the MSP432P401R operating at a clock rate of 48MHz . Due to variations in models, background noise, audio quality, and other factors the actual cycles an application requires may differ from these estimates.

ConditionAverage MCPS
(% of Available Cycles)		Maximum MCPS
(% of Available Cycles)
Background Mode2
(4.3%)		2
(4.3%)
Active Search Mode
(1 model)		6.2
(13%)		7.7
(16%)
Active Search Mode
(5 models)		17.3
(36%)		22.1
(46%)

Power Consumption
The MinHMM speech recognizer consumes arrays of sampled audio data. Typically the arrays consist of 160 samples collected at an 8kHz sample rate, so an array of data is provided to MinHMM each 20ms. The power needed to process these arrays is based on the cycles required for processing the data arrays, which was discussed in the prior paragraph.  The total power consumption of an application will in addition depend heavily on hardware and application design. For lowest power operation the design should utilize low-power best practices. The hardware should use low-power components for audio collection, such as the microphone and pre-amplifier. The application should shut down peripherals during periods when they are not being used and use the lowest clocking rates necessary. Application software should utilize the low-power modes of the processor, such as LPM3, during idle times. Excluding the external hardware, estimates of processor power consumption for an application utilizing these practices ranges from 250µA in background mode to 1.8mA in active search mode with five models active.

The current MinHMM example program demonstrates the use of the MinHMM speech recognizer, but for simplicity of illustration it was not designed for lowest power operation. For example, during idle times it uses LPM0 and the 48MHz clock remains running. 

Validation Information

This release is an EA version.

Versioning

This is build 0_90_00_00 of  Speech Recognizer Project.

Technical Support and Product Updates

Additional Resources

For more information, visit www.ti.com.