NOISE REDUCTION UNIT (NRU) is designed for using with
low bit rate vocoders 600-4400 bps. NRU makes noticeable positive effect
for vocoder speech depending from noise type for input signal-to-noise
ratios (SNR) 0…+20 dB.
Separation of speech and noises with overlapping spectrums is possible
on the base of adaptive nonlinear processing in the frequency domain.
The main discriminate sign between noise and speech is the time of spectrum
stationarity. Suggesting that noise spectrum characteristics are changed
slowly than speech spectrum characteristics.
Standard ITU-T P.862 PESQ-MOS criteria is used for output speech quality
measurements
The maximum advantage of using proposed noise reduction algorithm is up
to 6-8 dB of input SNR depending of noise type (car noise, street noise).
Noise reduction algorithm is implemented as fixed point
C model for PC. It may be implemented in real-time on the base of TI's
TMS320c54x/c55x DSP families under contract with resources no more than
20 MIPS and 10 Kwords of memory.
Following speech files demonstrate the quality of NRU with RMELP vocoder
4400 bps:
Original speech |
|
Car noise, SNR +12 dB
Speech with noise
|
|
Vocoder 4400 bps without NRU |
|
NRU output |
|
Vocoder 4400 bps with NRU |
|
Car noise, SNR +9 dB
Speech with noise |
|
Vocoder 4400 bps without NRU |
|
NRU output |
|
Vocoder 4400 bps with NRU |
|
Car noise, SNR +6 dB
Speech with noise |
|
Vocoder 4400 bps without NRU |
|
NRU output |
|
Vocoder 4400 bps with NRU |
|
Car noise, SNR +3 dB
Speech with noise |
|
Vocoder 4400 bps without NRU |
|
NRU output |
|
Vocoder 4400 bps with NRU |
|
|
Description
Implementation
Demonstration
All technical and commercial questions about products, prices,
terms of delivery, software libraries, source code etc. please send us by e-mail:
|