mglsadf

Functions

int main(int argc, char *argv[])

mglsadf [ option ] mgcfile [ infile ]

  • -m int

    • order of coefficients \((0 \le M)\)

  • -a double

    • all-pass constant \((|\alpha| < 1)\)

  • -c int

    • gamma \(\gamma = -1 / C\) \((1 \le C)\)

  • -p int

    • frame period \((1 \le P)\)

  • -i int

    • interpolation period \((0 \le I \le P/2)\)

  • -P int

    • order of Pade approximation \((4 \le L \le 7)\)

  • -t bool

    • transpose filter

  • -k bool

    • filtering without gain

  • mgcfile str

    • double-type mel-generalized cepstral coefficients

  • infile str

    • double-type input sequence

  • stdout

    • double-type output sequence

In the below example, an exciation signal generated from pitch information is passed through the MLSA filter built from mel-cepstral coefficients data.mcep.

excite < data.pitch | mglsadf data.mcep > data.syn

Parameters

  • argc – [in] Number of arguments.

  • argv – [in] Argument vector.

Returns

0 on success, 1 on failure.

See also

imglsadf mgcep

class sptk::MlsaDigitalFilter

Apply MLSA digital filter for speech synthesis to signals.

The transfer function \(H(z)\) of an MLSA filter is

\[\begin{split}\begin{eqnarray} H(z) &=& \exp \sum_{m=0}^M b(m) \varPhi_m(z) \\ &\equiv& K \cdot D(z) \end{eqnarray}\end{split}\]
where
\[\begin{split} \varPhi_m(z) = \left\{ \begin{array}{ll} 1, & m = 0 \\ \dfrac{(1-\alpha^2)z^{-1}}{1-\alpha z^{-1}} \tilde{z}^{-(m-1)}, & m \ge 1 \end{array} \right. \end{split}\]
and
\[\begin{split}\begin{eqnarray} K &=& \exp \, b(0), \\ D(z) &=& \exp \sum_{m=1}^M b(m) \varPhi_m(z). \end{eqnarray}\end{split}\]
The filter \(D(z)\) can be implemented by approximating exponential function as an \(L\)-th order rational function \(R_L(\cdot)\) using the modified Pade approximation:
\[ D(z) = \exp F(z) \simeq R_L(F(z)). \]
The block diagram of the basic filter \(F(z)\) is shown as below.

../_images/mglsadf_1.png

To improve the accuracy of the approximation, the basic filter is decomposed as

\[ F(z) = F_1(z) + F_2(z) \]
where
\[\begin{split}\begin{eqnarray} F_1(z) &=& b(1) \varPhi_1(z), \\ F_2(z) &=& \exp \sum_{m=2}^M b(m) \varPhi_m(z). \end{eqnarray}\end{split}\]

Given the \(M\)-th order MLSA filter coefficients,

\[ \begin{array}{cccc} b(0), & b(1), & \ldots, & b(M), \end{array} \]
an output signal is obtained by applying \(H(z)\) to an input signal in time domain.

Public Functions

MlsaDigitalFilter(int num_filter_order, int num_pade_order, double alpha, bool transposition)
Parameters

  • num_filter_order – [in] Order of filter coefficients, \(M\).

  • num_pade_order – [in] Order of Pade approximation, \(L\).

  • alpha – [in] All-pass constant, \(\alpha\).

  • transposition – [in] If true, use transposed form filter.

inline int GetNumFilterOrder() const
Returns

Order of coefficients.

inline int GetNumPadeOrder() const
Returns

Order of Pade approximation.

inline double GetAlpha() const
Returns

All-pass constant.

inline bool GetTranspositionFlag() const
Returns

True if transposed form is used.

inline bool IsValid() const
Returns

True if this object is valid.

bool Run(const std::vector<double> &filter_coefficients, double filter_input, double *filter_output, MlsaDigitalFilter::Buffer *buffer) const
Parameters

  • filter_coefficients – [in] \(M\)-th order MLSA filter coefficients.

  • filter_input – [in] Input signal.

  • filter_output – [out] Output signal.

  • buffer – [inout] Buffer.

Returns

True on success, false on failure.

bool Run(const std::vector<double> &filter_coefficients, double *input_and_output, MlsaDigitalFilter::Buffer *buffer) const
Parameters

  • filter_coefficients – [in] \(M\)-th order MLSA filter coefficients.

  • input_and_output – [inout] Input/output signal.

  • buffer – [inout] Buffer.

Returns

True on success, false on failure.

class Buffer

Buffer for MlsaDigitalFilter class.

class sptk::MglsaDigitalFilter

Apply MGLSA digital filter for speech synthesis to signals.

The transfer function \(H(z)\) of an MLSA filter is

\[\begin{split}\begin{eqnarray} H(z) &=& s_\gamma^{-1} \left( \sum_{m=0}^M b'_\gamma(m) \varPhi_m(z) \right) \\ &\equiv& K \cdot D(z) \end{eqnarray}\end{split}\]
where
\[\begin{split} \varPhi_m(z) = \left\{ \begin{array}{ll} 1, & m = 0 \\ \dfrac{(1-\alpha^2)z^{-1}}{1-\alpha z^{-1}} \tilde{z}^{-(m-1)}, & m \ge 1 \end{array} \right. \end{split}\]
and
\[\begin{split}\begin{eqnarray} K &=& s_\gamma^{-1} \left( b'_\gamma(0) \right), \\ D(z) &=& s_\gamma^{-1} \left(\sum_{m=1}^M b'_\gamma(m) \varPhi_m(z)\right). \end{eqnarray}\end{split}\]
The filter \(D(z)\) is implemented by cascading filter \(1/B(z)\):
\[\begin{split}\begin{eqnarray} D(z) &=& \left( \frac{1}{B(z)} \right)^C \\ &=& \left( \frac{1} {1 + \gamma \sum_{m=1}^M b'_\gamma(m) \varPhi_m(z)} \right)^C . \end{eqnarray}\end{split}\]

../_images/mglsadf_2.png

The filter coeffcients \(b'_\gamma(m)\) are obtained by applying gain normalization to the MLSA filter coefficients:

\[ \begin{array}{cccc} b(0), & b(1), & \ldots, & b(M). \end{array} \]
An output signal is obtained by applying \(H(z)\) to an input signal in time domain.

Public Functions

MglsaDigitalFilter(int num_filter_order, int num_pade_order, int num_stage, double alpha, bool transposition)
Parameters

  • num_filter_order – [in] Order of filter coefficients, \(M\).

  • num_pade_order – [in] Order of Pade approximation, \(L\).

  • num_stage – [in] Number of stages, \(C\). If zero, MLSA filter is used.

  • alpha – [in] All-pass constant, \(\alpha\).

  • transposition – [in] If true, use transposed form filter.

inline int GetNumFilterOrder() const
Returns

Order of coefficients.

inline int GetNumPadeOrder() const
Returns

Order of Pade approximation.

inline int GetNumStage() const
Returns

Number of stages.

inline double GetAlpha() const
Returns

All-pass constant.

inline bool GetTranspositionFlag() const
Returns

True if transposed form is used.

inline bool IsValid() const
Returns

True if this object is valid.

bool Run(const std::vector<double> &filter_coefficients, double filter_input, double *filter_output, MglsaDigitalFilter::Buffer *buffer) const
Parameters

  • filter_coefficients – [in] \(M\)-th order MGLSA filter coefficients.

  • filter_input – [in] Input signal.

  • filter_output – [out] Output signal.

  • buffer – [inout] Buffer.

Returns

True on success, false on failure.

bool Run(const std::vector<double> &filter_coefficients, double *input_and_output, MglsaDigitalFilter::Buffer *buffer) const
Parameters

  • filter_coefficients – [in] \(M\)-th order MGLSA filter coefficients.

  • input_and_output – [inout] Input/output signal.

  • buffer – [inout] Buffer.

Returns

True on success, false on failure.

class Buffer

Buffer for MglsaDigitalFilter class.