SPEAKING PARTNER: AN ARM7BASED MULTIMEDIA HANDHELD DEVICE Wonyong Sung, Hoseok Chang, Wonchul Lee, S

1
SPEAKING PARTNER AN ARM7-BASED MULTIMEDIA
HANDHELD DEVICEWonyong Sung, Hoseok Chang,
Wonchul Lee, Suhong Ryu

• School of Electrical Engineering
• Seoul National University
• Wonyong Sung
• Oct. 17, 2002

2
Contents

• Introduction
• Software optimization techniques
• Implementation of MP3 Audio decoder
• Implementation of speech recognition
• Power consumption of the system
• Conclusion

3
Introduction

• Speaking Partner
• Handheld educational device for kids
• Animation, MP3 playback, speech recognition

2,9 240160 black and white 16-gray STN LCD
128Kbyte, System ROM
Or Smart Media Card(SMC) Store programs
contents, Used as a solid-state hard-disk
60 MHz, 8KB unified cache memory
4
Software Implementation

• Implementation
• MP3 audio decoding
• Speech recognition
• Animation(LZW image compression)
• Using a simple real-time operating system

Tasks
Top priority jobs
5
Software Optimization(1)

• Based on 32bit RISC processor
• ARM7TDMI
• 328 bits multiplier
• No support DSP specific features
• Hardware loop control, automatic address
  generation, multiple buses
• Need to optimize the computation intensive
  algorithms
• MP3 decoding, speech recognition
• Low level optimization
• Using ARM CPU specific features

6
Software optimization(2)

• ARM7TDMI
• A fairy large number of registers(31)
• Reduce the memory access, good compiler
  performance
• Conditional execution
• Reduce control overhead
• 32bit barrel shifter
• Execute shift/rotation with ALU operations
• Block load/store transfer instructions
• Need to convert the floating-point version to
  fixed-point one
• Using automatic scaling method
• Optimization
• Algorithm optimization
• Loop transformation
• Using ARM CPU specific features

7
Implementation of MP3 Decoder

• MPEG 1/2 Layer-III(MP3) algorithm
• Subband synthesis, IMDCT, requantization
• gt 85 of processing time
• Huffman decoding, Unpacking scale-factor
• gt control intensive part
• Inverse Modified Discrete Cosine Transform(IMDCT)
• Britanak Rao algorithm
• Using low level optimization
• Block transfer instructions(LDM/STM)

8
IMDCT

• Reducing multiplier accuracy vs. using block
  transfer
• A -gt B 17.8
• B -gt C 8.1

SNR 3232 bits 91.61dB 3216 bits 82.45dB
The effect of reducing the data memory gt the
effect of reducing the precision for multiply
9
Implementation of Speech recognition

• Connected word speaker independent

8KHz(12bits), 12th order LPCC
Pattern recognizer, Hidden Markov Model
Viterbi algorithm Beam-pruning gt reduce total
execution time about 60
Recognition unit word Small vocabulary size
Examine just the previous word Reduced the
latency of recognition, small storage
10
Profiling results

• Using software optimization methods
• Loop transformations
• Post increment/decrement conversion
• 16 bit multiplications
• Test
• 215 word-set
• Over 90 recognition rate

Execution time gt Reduced about 30 45
11
Power Consumption

• The system is operating using two AA-size 1.5V
  batteries
• Most digital analog circuits 3.3V
• CPU core 2.5V, LCD 21V, Audio amp 150mW

Current consumption according to the activity
Current consumption at each H/W block
12
Conclusion

• MP3 decoding, speech recognition, animation
• Implemented in software using ARM7TDMI
• Lacks of features in programmable digital signal
  processors
• Need software optimization techniques
• The effect of reducing the data transfer
• gt the effect of reducing the precision for
  multiply
• Power analysis
• Current consumed at the DRAM
• Current consumed at the CPU core
• gt reducing cache miss is most important for
  lowering power consumption