o ZhlY@sdZddlmZmZmZmZmZddlZddl m Z m Z m Z m Z ddlmZddlmZddlmZmZmZeeZGd d d eZd gZdS) z)Feature extractor class for UnivNetModel.)AnyDictListOptionalUnionN)mel_filter_bankoptimal_fft_length spectrogramwindow_function)SequenceFeatureExtractor) BatchFeature)PaddingStrategy TensorTypeloggingc)seZdZdZgdZ              dFdedededededededede ededede ed ed!ed"ed#ed$ed%ed&ed'ef(fd(d) Z d*d+Z d,d-Z d.e jd/e jfd0d1Z dGd2ed3e e jjd/e jfd4d5ZdGd/ee jfd6d7Z     dHd8ee jeeee jeeefde ed9eeeefd:e ed;eded?e ede ed@e edAe eeefd/efdBdCZd/eeefffdDdE ZZS)IUnivNetFeatureExtractora Constructs a UnivNet feature extractor. This class extracts log-mel-filter bank features from raw speech using the short time Fourier Transform (STFT). The STFT implementation follows that of TacoTron 2 and Hifi-GAN. This feature extractor inherits from [`~feature_extraction_sequence_utils.SequenceFeatureExtractor`] which contains most of the main methods. Users should refer to this superclass for more information regarding those methods. Args: feature_size (`int`, *optional*, defaults to 1): The feature dimension of the extracted features. sampling_rate (`int`, *optional*, defaults to 24000): The sampling rate at which the audio files should be digitalized expressed in hertz (Hz). padding_value (`float`, *optional*, defaults to 0.0): The value to pad with when applying the padding strategy defined by the `padding` argument to [`UnivNetFeatureExtractor.__call__`]. Should correspond to audio silence. The `pad_end` argument to `__call__` will also use this padding value. do_normalize (`bool`, *optional*, defaults to `False`): Whether to perform Tacotron 2 normalization on the input. Normalizing can help to significantly improve the performance for some models. num_mel_bins (`int`, *optional*, defaults to 100): The number of mel-frequency bins in the extracted spectrogram features. This should match `UnivNetModel.config.num_mel_bins`. hop_length (`int`, *optional*, defaults to 256): The direct number of samples between sliding windows. Otherwise referred to as "shift" in many papers. Note that this is different from other audio feature extractors such as [`SpeechT5FeatureExtractor`] which take the `hop_length` in ms. win_length (`int`, *optional*, defaults to 1024): The direct number of samples for each sliding window. Note that this is different from other audio feature extractors such as [`SpeechT5FeatureExtractor`] which take the `win_length` in ms. win_function (`str`, *optional*, defaults to `"hann_window"`): Name for the window function used for windowing, must be accessible via `torch.{win_function}` filter_length (`int`, *optional*, defaults to 1024): The number of FFT components to use. If `None`, this is determined using `transformers.audio_utils.optimal_fft_length`. max_length_s (`int`, *optional*, defaults to 10): The maximum input length of the model in seconds. This is used to pad the audio. fmin (`float`, *optional*, defaults to 0.0): Minimum mel frequency in Hz. fmax (`float`, *optional*): Maximum mel frequency in Hz. If not set, defaults to `sampling_rate / 2`. mel_floor (`float`, *optional*, defaults to 1e-09): Minimum value of mel frequency banks. Note that the way [`UnivNetFeatureExtractor`] uses `mel_floor` is different than in [`transformers.audio_utils.spectrogram`]. center (`bool`, *optional*, defaults to `False`): Whether to pad the waveform so that frame `t` is centered around time `t * hop_length`. If `False`, frame `t` will start at time `t * hop_length`. compression_factor (`float`, *optional*, defaults to 1.0): The multiplicative compression factor for dynamic range compression during spectral normalization. compression_clip_val (`float`, *optional*, defaults to 1e-05): The clip value applied to the waveform before applying dynamic range compression during spectral normalization. normalize_min (`float`, *optional*, defaults to -11.512925148010254): The min value used for Tacotron 2-style linear normalization. The default is the original value from the Tacotron 2 implementation. normalize_max (`float`, *optional*, defaults to 2.3143386840820312): The max value used for Tacotron 2-style linear normalization. The default is the original value from the Tacotron 2 implementation. model_in_channels (`int`, *optional*, defaults to 64): The number of input channels to the [`UnivNetModel`] model. This should match `UnivNetModel.config.model_in_channels`. pad_end_length (`int`, *optional*, defaults to 10): If padding the end of each waveform, the number of spectrogram frames worth of samples to append. The number of appended samples will be `pad_end_length * hop_length`. return_attention_mask (`bool`, *optional*, defaults to `True`): Whether or not [`~UnivNetFeatureExtractor.__call__`] should return `attention_mask`. )input_featuresnoise_sequence padding_mask]Fd hann_window N& .>?h㈵> 'ă@@T feature_size sampling_rate padding_value do_normalize num_mel_bins hop_length win_length win_function filter_length max_length_sfminfmax mel_floorcentercompression_factorcompression_clip_val normalize_min normalize_maxmodel_in_channelspad_end_lengthc s tjd||||d|||_||_||_||_||_| |_| |_| dur-t |d} | |_ | |_ | |_ | ||_ |jdurGt|j|_n|j|_|jdd|_t|j|jdd|_t|j|j|j|j |jddd|_||_||_||_||_||_||_||_dS) N)r#r$r%return_attention_maskrT)Z window_lengthnameZperiodicZslaney)Znum_frequency_binsZnum_mel_filtersZ min_frequencyZ max_frequencyr$ZnormZ mel_scale)super__init__r&r'r(r)r*r+r-floatr.r/r,num_max_samplesr n_fftn_freqsr windowrr$ mel_filtersr0r1r2r3r4r5r6)selfr#r$r%r&r'r(r)r*r+r,r-r.r/r0r1r2r3r4r5r6r7kwargs __class__r:e/var/www/auris/lib/python3.10/site-packages/transformers/models/univnet/feature_extraction_univnet.pyr<esT     z UnivNetFeatureExtractor.__init__cCsd||j|j|jdS)Nr8rr3r4rCr r:r:rG normalizez!UnivNetFeatureExtractor.normalizecCs|j|j|j|ddS)Nrr8rHrIr:r:rG denormalizerKz#UnivNetFeatureExtractor.denormalizewaveformreturnc Cstj|t|j|jdt|j|jdfdd}t||j|j|j|jd|jddd }tt |dt |d|j }t |j j|}ttj||jdd|j}|jS)a Calculates log MEL spectrograms from a batch of waveforms. Note that the input waveform(s) will be padded by `int(self.n_fft - self.hop_length) / 2` on both sides using the `reflect` padding mode. Args: waveform (`np.ndarray` of shape `(length,)`): The input waveform. This must be a single real-valued, mono waveform. Returns: `numpy.ndarray`: Array containing a log-mel spectrogram of shape `(num_frames, num_mel_bins)`. r8Zreflect)modeN)rAZ frame_lengthr(Z fft_lengthpowerr0rBr/)Za_minZa_max)nppadintr?r(r rAr0sqrtrealimagr/matmulrBTlogZclipr2r1)rCrMZcomplex_spectrogramZamplitude_spectrogrammel_spectrogramZlog_mel_spectrogramr:r:rGrZs0& z'UnivNetFeatureExtractor.mel_spectrogram noise_length generatorcCs0|dur tj}||jf}|j|tjd}|S)a Generates a random noise sequence of standard Gaussian noise for use in the `noise_sequence` argument of [`UnivNetModel.forward`]. Args: spectrogram_length (`int`): The length (dim 0) of the generated noise. model_in_channels (`int`, *optional*, defaults to `None`): The number of features (dim 1) of the generated noise. This should correspond to the `model_in_channels` of the [`UnivNetGan`] model. If not set, this will default to `self.config.model_in_channels`. generator (`numpy.random.Generator`, *optional*, defaults to `None`) An optional `numpy.random.Generator` random number generator to control noise generation. If not set, a new generator with fresh entropy will be created. Returns: `numpy.ndarray`: Array containing random standard Gaussian noise of shape `(noise_length, model_in_channels)`. Ndtype)rQrandomZ default_rngr5Zstandard_normalfloat32)rCr[r\Z noise_shapenoiser:r:rGgenerate_noises   z&UnivNetFeatureExtractor.generate_noisecs0dd|D}durfddt|D}|S)a Removes padding from generated audio after running [`UnivNetModel.forward`]. This returns a ragged list of 1D audio waveform arrays and not a single tensor/array because in general the waveforms will have different lengths after removing padding. Args: waveforms (`torch.FloatTensor` of shape `(batch_size, sequence_length)`): The batched output waveforms from the [`UnivNetModel`]. waveform_lengths (`torch.FloatTensor` of shape `(batch_size,)`, *optional*): The batched lengths of each waveform before padding. Returns: `List[np.ndarray]`: A ragged list of 1D waveform arrays with padding removed. cSs"g|] }|jdddqS)cpuT)Zdevicecopy)detachtonumpy.0rMr:r:rG s"z8UnivNetFeatureExtractor.batch_decode..Ncs g|] \}}|d|qSNr:)riirMwaveform_lengthsr:rGrjs ) enumerate)rCZ waveformsrnr:rmrG batch_decodesz$UnivNetFeatureExtractor.batch_decode raw_speechpadding max_length truncationpad_to_multiple_of return_noisepad_end pad_lengthr7return_tensorsc sl| dur| nj} |dur(|jkr'tdjjdjdjd|d n tdjjdt|tj o?t |j d k}|rPt |j d krPtd |pct|t t foct|d tj t t f}|rnd d|D}n&|st|tj stj|tjd}nt|tj r|jttjur|tj}|stj|tjdg}| rdurnjfdd|D}td|i}j|||dur|nj||| d}|d}fdd|D}t|d trdd|D|d<n dd|D|d<|d}|durdd|D|d<|rfdd|dD}||d<| r*fdd|dD|d<| dur4|| }|S)a Main method to featurize and prepare for the model one or several sequence(s). Args: raw_speech (`np.ndarray`, `List[float]`, `List[np.ndarray]`, `List[List[float]]`): The sequence or batch of sequences to be padded. Each sequence can be a numpy array, a list of float values, a list of numpy arrays or a list of list of float values. Must be mono channel audio, not stereo, i.e. single float per timestep. sampling_rate (`int`, *optional*): The sampling rate at which the `raw_speech` input was sampled. It is strongly recommended to pass `sampling_rate` at the forward call to prevent silent errors and allow automatic speech recognition pipeline. padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `True`): Select a strategy to pad the input `raw_speech` waveforms (according to the model's padding side and padding index) among: - `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single sequence if provided). - `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum acceptable input length for the model if that argument is not provided. - `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different lengths). If `pad_end = True`, that padding will occur before the `padding` strategy is applied. max_length (`int`, *optional*): Maximum length of the returned list and optionally padding length (see above). truncation (`bool`, *optional*, defaults to `True`): Activates truncation to cut input sequences longer than `max_length` to `max_length`. pad_to_multiple_of (`int`, *optional*): If set will pad the sequence to a multiple of the provided value. This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability `>= 7.5` (Volta), or on TPUs which benefit from having sequence lengths be a multiple of 128. return_noise (`bool`, *optional*, defaults to `True`): Whether to generate and return a noise waveform for use in [`UnivNetModel.forward`]. generator (`numpy.random.Generator`, *optional*, defaults to `None`): An optional `numpy.random.Generator` random number generator to use when generating noise. pad_end (`bool`, *optional*, defaults to `False`): Whether to pad the end of each waveform with silence. This can help reduce artifacts at the end of the generated audio sample; see https://github.com/seungwonpark/melgan/issues/8 for more details. This padding will be done before the padding strategy specified in `padding` is performed. pad_length (`int`, *optional*, defaults to `None`): If padding the end of each waveform, the length of the padding in spectrogram frames. If not set, this will default to `self.config.pad_end_length`. do_normalize (`bool`, *optional*): Whether to perform Tacotron 2 normalization on the input. Normalizing can help to significantly improve the performance for some models. If not set, this will default to `self.config.do_normalize`. return_attention_mask (`bool`, *optional*): Whether to return the attention mask. If left to the default, will return the attention mask according to the specific feature_extractor's default. [What are attention masks?](../glossary#attention-mask) return_tensors (`str` or [`~utils.TensorType`], *optional*): If set, will return tensors instead of list of python integers. Acceptable values are: - `'tf'`: Return TensorFlow `tf.constant` objects. - `'pt'`: Return PyTorch `torch.np.array` objects. - `'np'`: Return Numpy `np.ndarray` objects. Nz3The model corresponding to this feature extractor: z& was trained using a sampling rate of zI. Please make sure that the provided `raw_speech` input was sampled with z and not .zDIt is strongly recommended to pass the `sampling_rate` argument to `zN()`. Failing to do so can result in silent errors that might be hard to debug.rr8z2Only mono-channel audio is supported for input to rcSg|] }tj|tjdqSr]rQasarrayr`)riZspeechr:r:rGrjz4UnivNetFeatureExtractor.__call__..r]cs(g|]}tj|djfjdqS)r)Zconstant_values)rQrRr(r%rh)rxrCr:rGrjsr)rrrsrtrur7cg|]}|qSr:)rZrhrCr:rGrjscSr{r|r}riZmelr:r:rGrjrcSsg|]}|tjqSr:)astyperQr`rr:r:rGrjsattention_maskcSr{r|)rQr~Zint32)riarrayr:r:rGrjrrcsg|] }|jdqS)r)rbshaperir )r\rCr:rGrjsrcrr:)rJrrr:rGrjs )r&r$ ValueErrorrF__name__loggerwarning isinstancerQndarraylenrlisttupler~r`r^Zfloat64rr6r rRr>getrZconvert_to_tensors)rCrqr$rrrsrtrurvr\rwrxr&r7ryZis_batched_numpyZ is_batchedZbatched_speechZ padded_inputsrZmel_spectrogramsrrar:)r\rxrCrG__call__sL  "          z UnivNetFeatureExtractor.__call__cs.t}gd}|D] }||vr||=q |S)N)rArBr?r@r>)r;to_dict)rCoutputnamesr9rEr:rGrs zUnivNetFeatureExtractor.to_dict)rrrFrrrrrrrNrFrrr r!r"rTrk) NTNTNTNFNNNN)r __module__ __qualname____doc__Zmodel_input_namesrSr=boolstrrr<rJrLrQrrZr_ Generatorrbrrprrrr rrrr __classcell__r:r:rErGrsE     L3   "       ""r)rtypingrrrrrrgrQZ audio_utilsrr r r Z!feature_extraction_sequence_utilsr Zfeature_extraction_utilsr utilsrrrZ get_loggerrrr__all__r:r:r:rGs     1