o i@sdZddlZddlmZddlmZddlZddlZddl Z ddl m Z ddl mZdZ  ddejd ed ed ed ed edejfddZ   ddejd edededeedeedejfddZGdddeZdS)zGriffin-Lim related modules.N)partial)Optional)parse)check_argument_typesg|=lmspcfsn_fftn_melsfminfmaxreturnc Cs||jd|ks J|durdn|}|dur|dn|}td|}tjj|||||d}tj|}tt t ||j j S)a{Convert log Mel filterbank to linear spectrogram. Args: lmspc: Log Mel filterbank (T, n_mels). fs: Sampling frequency. n_fft: The number of FFT points. n_mels: The number of mel basis. f_min: Minimum frequency to analyze. f_max: Maximum frequency to analyze. Returns: Linear spectrogram (T, n_fft // 2 + 1). Nrg$@)srrr r r ) shapenppowerlibrosafiltersmellinalgpinvmaximumEPSdotT) rrrr r r mspc mel_basis inv_mel_basisrM/home/ubuntu/.local/lib/python3.10/site-packages/espnet2/utils/griffin_lim.py logmel2linears   r!hann spcn_shift win_lengthwindown_iterc Cs|jd|ddks Jttjtdkr2t|j}tj||||||jddkr,dndd}|St dt| tj j}t dtj tjj|j}tj|||||d }t|D]} t d ttj|||||d }tj|||||d }q_|S) aConvert linear spectrogram into waveform using Griffin-Lim. Args: spc: Linear spectrogram (T, n_fft // 2 + 1). n_fft: The number of FFT points. n_shift: Shift size in points. win_length: Window length in points. window: Window function type. n_iter: The number of iterations. Returns: Reconstructed waveform (N,). r rz0.7.0TF)Sr( hop_lengthr&r'centerzlibrosa version is old. use slow version of Grriffin-Lim algorithm.if you want to use fast Griffin-Lim, please update librosa via `source ./path.sh && pip install librosa==0.7.0`.y@)r'y?)rVr __version__rabsr griffinlimloggingwarningastypecomplexexppirandomrandistftrangeanglestft) r$rr%r&r'r(ycspcanglesirrr griffin_lim6s4  r@c@steZdZdZ       ddedededed ed eed ed ed eefddZddZde j de j fddZ dS)Spectrogram2Waveformz*Spectrogram to waveform conversion module.Nr"rr%rr r&r'r r griffin_lim_itersc CstsJ||_|durtt|||||dnd|_tt||||| d|_t||||| d|_|dur>|jj||||ddSdS)aInitialize module. Args: fs: Sampling frequency. n_fft: The number of FFT points. n_shift: Shift size in points. n_mels: The number of mel basis. win_length: Window length in points. window: Window function type. f_min: Minimum frequency to analyze. f_max: Maximum frequency to analyze. griffin_lim_iters: The number of iterations. N)rrr r r )rr%r&r'r()rr r r )rrrr!r@dictparamsupdate) selfrr%rr r&r'r r rCrrr __init__rs4  zSpectrogram2Waveform.__init__cCsB|jjd}|jD]\}}||d|d7}q |d7}|S)N(=z, )) __class____name__rEitems)rGretvalkvrrr __repr__s zSpectrogram2Waveform.__repr__r$r cCsJ|j}|j}|}|jdur||}||}t|j||dS)zConvert spectrogram to waveform. Args: spc: Log Mel filterbank (T_feats, n_mels) or linear spectrogram (T_feats, n_fft // 2 + 1). Returns: Tensor: Reconstructed waveform (T_wav,). N)devicedtype) rSrTcpunumpyr!r@torchtensorto)rGr$rSrTwavrrr __call__s     zSpectrogram2Waveform.__call__)NNNr"NNrB) rM __module__ __qualname____doc__intrstrrHrRrWTensorr[rrrr rAos<  5rA)NN)Nr"r#)r^r0 functoolsrtypingrrrVrrWpackaging.versionrr, typeguardrrndarrayr_r!r`r@objectrArrrr s\     % 9