Post on 19-Jan-2016
SPEECH CODING
Maryam ZebarjadAlessandro Chiumento
Supervisor : Sylwester Szczpaniak
Outline
Properties of speech signals Why coding ? Implemented tecniques
Differential Pulse-Code Modulation DCT Tranfrorm Coder LPC Vocoder
Results
SPEECH PROPERTIES
Speech is produced when air is forced from the lungs through the vocal cords and along the vocal tract.
It can be modeled by two states:
Voiced Speech: - produced by the vibrations of the vocal cords.- quasi-periodic in the time domain and harmonically
structured in the frequency domain.
Unvoiced Speech: - produced , for example, by high speed air passing through a constriction in the vocal tract (mouth and
lips)- random-like and broadband (like white noise).
Why coding ?
Original speech signal has to be processed in order to be :
MINIMIZE DIMENSIONS (storage)
MINIMIZE BITRATE (transmission)
VOIPMOBILE TELEPHONY
DPCM
We have done DPCM about a wave file and here is the result for different prediction orders:
-we have the coder and decoder signal for the prediction orders of 1, 2, 5, 10, 19.
-we have corresponding wave files for each stage-we also have the SNR for each prediction order
For the auto correlation method these were the basic formula as previously stated
The DPCM Method with autocorrelation
The Sriginal Signal
Coder Signal for Prediction Order of 1
Decoder Signal for Prediction Order of 1
Coder Signal for the Prediction order of 2
Decoder Signal for the Prediction Order of 2
Decoder Signal for the Prediction Order of 2
Coder Signal for the Prediction Order of 5
Decoder Signal for the Prediction Order of 5
Coder Signal for the Prediction Order of 10
Decoder Signal for the Prediction Order of 10
Coder Signal for the Prediction Order of 19
Decoder Signal for the Prediction Order of 19
SNR
Then by the following formula we calculate the Decoder SNR for each prediction order
LPC VocoderVocoders rely strongly on the properties of speech.
Two – state excitation model: - pulses for voiced signal- random noise for unvoiced
signalVocal tract is modeled as an all-pole function.
Source-System synthesis model
where
LPC Vocoder
We have to find: - pitch period- gain- poles of the system
LPC Vocoder
V/UV DETECTION is done by taking the energy of each frame and compare it to a threshold. Taking the zero-crossing rate and compare it to a threshold.
PITCH DETECTION is done by Autocorrelation method : we cross-correlate the signal with it self,
the output has a max after the pitch period.
POLES OF THE SYSTEM are estimated using: LPC, in our case the LEVINSON-DURBIN algorithm
GAIN IS ESTIMATED : If the frame is UnVoiced we take the sqrt of the average power of
the frame. If the frame is Voiced we use the average power for every pitch
period.
LPC VocoderORIGINAL SAMPLE
SYNTHETIZED SAMPLES
DCT Transform Coder
There is no standard Same structure than vocoder
DCT Transform Coder
Discrete Cosine Trasform is a unitary transform that expresses the incoming signal as a finite sum of cosine functions:
So if the signal is periodic we need a “small” number of cosines (coefficients)insteadif the signal is non periodic the cosines have to be many more.
DCT Transform CoderVoiced frame : waveform DCT coefficients
Unvoiced frame : waveform DCT coefficients
DCT Transform CoderORIGINAL SAMPLE
Synthetized sample 22.5ms720 coeff V1460 coeff UV
22.5ms40 coeff V1460 coeff UV
50ms720 coeff V1460 coeff UV