Speech recognition frontend on Cell BE Pavel Bazika - - PowerPoint PPT Presentation

▶

Dec 18, 2022 386 likes •453 views

IBM - CVUT Student Research Projects Speech recognition frontend on Cell BE Pavel Bazika (bazikp1@fel.cvut.cz) Speech recognizer Input speech is represented by samples Inner format is 25ms length frames FRONTEND speech comparison

SLIDE 1

IBM - CVUT Student Research Projects

Speech recognition frontend

n Cell BE

Pavel Bazika (bazikp1@fel.cvut.cz)

SLIDE 2 IBM - CVUT Student Research Projects 2

Speech recognizer

speech FRONTEND

preprocessing
feature extraction

comparison vocabulary

word probability

Input speech is represented by samples
Inner format is 25ms length frames

SLIDE 3 IBM - CVUT Student Research Projects 3

Algorithms needed for speech recognition

Mean value subtraction
Preemphasis
Hamming window selection
FFT
Logarithm
Triangular filters
DCT

} cepstrum

SLIDE 4 IBM - CVUT Student Research Projects 4

Speed of our algorithm

Four frames are computed at once
Cepstrum calculation of 25 ms length frame

for input sampling frequency 8 kHz takes 3,7 μs

One SPU can process

2700 speeches in realtime

SLIDE 5 IBM - CVUT Student Research Projects 5

Cepstrum calculation comparison with Pentium 4

200 400 600 800 1000 1200 5000 10000 15000 20000 25000 30000 SPU F4S Pentium 4 Frame size Time [ns]

SLIDE 6 IBM - CVUT Student Research Projects 6

Highlights

Optimized algorithms for SPU, dual-issue

used when possible

FFT for four streams of data implemented
Pentium 4 is slower in every algorithm
Faster FFT than FFTW with SSE2 enabled
Input samples are converted to inner format

in parallel with mean value computation