U %d>=@s&ddlZddlmZddlmZmZmZmZmZm Z ddl Z ddl m Z ddl m Z ddlmZddlmZmZddlmZmZd d lmZd d lmZgZd ZGd ddejZGdddejZGddde jj ej!Z"Gddde jj ej!Z#GdddZ$GdddZ%eGdddZ&eGdddZ'GdddZ(eGdd d e'e&e$eZ)eGd!d"d"e'e&e%eZ*eGd#d$d$e(e&e$eZ+eGd%d&d&e(e&e%eZ,e+d'ej-d(d)d*Z.d+e._/e,d,ej-d-d)d*Z0d.e0_/e)d/ej-d(d)d0e1d1Z2d2e2_/e*d3ej-d-d)d0e1d1Z3d4e3_/dS)5N) dataclass)AnyDictListOptionalTupleUnion)Tensor)load_state_dict_from_url)mu_law_decoding) Tacotron2WaveRNN) GriffinLimInverseMelScale)utils)Tacotron2TTSBundlez.https://download.pytorch.org/torchaudio/modelscsLeZdZfddZeddZeeeefe e e fdddZ Z S)_EnglishCharProcessorcs.tt|_ddt|jD|_dS)NcSsi|]\}}||qSr).0isrrB/tmp/pip-unpacked-wheel-lbdmvq91/torchaudio/pipelines/_tts/impl.py sz2_EnglishCharProcessor.__init__..)super__init__rZ _get_chars_tokens enumerate_mappingself __class__rrrs  z_EnglishCharProcessor.__init__cCs|jSNrrrrrtokenssz_EnglishCharProcessor.tokenstextsreturncs,t|tr|g}fdd|D}t|S)Ncs"g|]}fdd|DqS)cs g|]}|jkrj|qSrr)rcrrr &s z=_EnglishCharProcessor.__call__...)lower)rtrrrr+&sz2_EnglishCharProcessor.__call__..) isinstancestrr _to_tensor)r r'indicesrrr__call__#s z_EnglishCharProcessor.__call__ __name__ __module__ __qualname__rpropertyr%rr/rrr r2 __classcell__rrr!rrs  rcsReZdZddfdd ZeddZeeeefe e e fddd Z Z S) _EnglishPhoneProcessorN dl_kwargscsDtt|_ddt|jD|_tjd|d|_d|_ dS)NcSsi|]\}}||qSrr)rrprrrr.sz3_EnglishPhoneProcessor.__init__..zen_us_cmudict_forward.ptr:z(\[[A-Z]+?\]|[_!'(),.:;? -])) rrrZ _get_phonesrrrZ_load_phonemizer _phonemizer_patternr r;r!rrr+s   z_EnglishPhoneProcessor.__init__cCs|jSr#r$rrrrr%2sz_EnglishPhoneProcessor.tokensr&csbt|tr|g}g}j|ddD]4}ddtj|D}|fdd|Dq"t|S)Nen_us)langcSsg|]}tdd|qS)z[\[\]])resub)rrrrrr+=sz3_EnglishPhoneProcessor.__call__..csg|]}j|qSrr))rr<rrrr+>s) r.r/r=rCfindallr>appendrr0)r r'r1Zphonesretrrrr26s z_EnglishPhoneProcessor.__call__r3rrr!rr9*s r9cs@eZdZd eeedfdd ZeddZd dd Z Z S) _WaveRNNVocoder)model min_level_dbcs td|_||_||_dS)N"V)rr _sample_rate_model _min_level_db)r rKrLr!rrrHs z_WaveRNNVocoder.__init__cCs|jSr#rNrrrr sample_rateNsz_WaveRNNVocoder.sample_rateNcCst|}dttj|dd}|jdk rL|j||j}tj|ddd}|j||\}}t||jj }t ||jj }| d}||fS)Ngh㈵>)minrr)rTmax) torchexplog10clamprPrOZinferrZ_unnormalize_waveformZn_bitsr Z n_classesZsqueeze)r mel_speclengthsZwaveformrrrforwardRs   z_WaveRNNVocoder.forward)rJ)N) r4r5r6r rfloatrr7rRr\r8rrr!rrIGs rIcs2eZdZfddZeddZdddZZS) _GriffinLimVocoderc s@td|_tdd|jddddd|_tdd d dd |_dS) NrMiPgg@@Zslaney)Zn_stftZn_melsrRZf_minZf_maxZ mel_scaleZnormir)Zn_fftpowerZ hop_lengthZ win_length)rrrNrrR_inv_melr _griffin_limrr!rrr`s"  z_GriffinLimVocoder.__init__cCs|jSr#rQrrrrrRssz_GriffinLimVocoder.sample_rateNcCsFt|}|d}||}|d}||}||fS)NTF)rVrWclonedetachZrequires_grad_rbrc)r rZr[specZ waveformsrrrr\ws    z_GriffinLimVocoder.forward)N)r4r5r6rr7rRr\r8rrr!rr^_s  r^c@seZdZejdddZdS) _CharMixinr(cCstSr#)rrrrrget_text_processorsz_CharMixin.get_text_processorNr4r5r6r TextProcessorrirrrrrgsrgc@s"eZdZddejdddZdS) _PhoneMixinNr:rhcCs t|dSNr:)r9r?rrrrisz_PhoneMixin.get_text_processorrjrrrrrlsrlc@s:eZdZUeed<eeefed<ddedddZdS)_Tacotron2Mixin_tacotron2_path_tacotron2_paramsNr:rhcCsNtf|j}td|j}|dkr(in|}t|f|}||||SN/)r rp _BASE_URLror load_state_dictevalr r;rKurlZ state_dictrrr get_tacotron2s   z_Tacotron2Mixin.get_tacotron2) r4r5r6r/__annotations__rrr rxrrrrrns rnc@sJeZdZUeeed<eeeefed<ddddZddddZ dS) _WaveRNNMixin _wavernn_path_wavernn_paramsNr:cCs|j|d}t|Srm) _get_wavernnrI)r r;Zwavernnrrr get_vocoders z_WaveRNNMixin.get_vocodercCsNtf|j}td|j}|dkr(in|}t|f|}||||Srq)r r|rsr{r rtrurvrrrr}s   z_WaveRNNMixin._get_wavernn) r4r5r6rr/ryrrr~r}rrrrrzs  rzc@seZdZddZdS)_GriffinLimMixincKstSr#)r^)r _rrrr~sz_GriffinLimMixin.get_vocoderN)r4r5r6r~rrrrrsrc@s eZdZdS)_Tacotron2WaveRNNCharBundleNr4r5r6rrrrrsrc@s eZdZdS)_Tacotron2WaveRNNPhoneBundleNrrrrrrsrc@s eZdZdS)_Tacotron2GriffinLimCharBundleNrrrrrrsrc@s eZdZdS)_Tacotron2GriffinLimPhoneBundleNrrrrrrsrz5tacotron2_english_characters_1500_epochs_ljspeech.pth&)Z n_symbols)rorpaLCharacter-based TTS pipeline with :py:class:`torchaudio.models.Tacotron2` and :py:class:`torchaudio.transforms.GriffinLim`. The text processor encodes the input texts character-by-character. Tacotron2 was trained on *LJSpeech* [:footcite:`ljspeech17`] for 1,500 epochs. You can find the training script `here `__. The default parameters were used. The vocoder is based on :py:class:`torchaudio.transforms.GriffinLim`. Please refer to :func:`torchaudio.pipelines.Tacotron2TTSBundle` for the usage. Example - "Hello world! T T S stands for Text to Speech!" .. image:: https://download.pytorch.org/torchaudio/doc-assets/TACOTRON2_GRIFFINLIM_CHAR_LJSPEECH.png :alt: Spectrogram generated by Tacotron2 .. raw:: html Example - "The examination and testimony of the experts enabled the Commission to conclude that five shots may have been fired," .. image:: https://download.pytorch.org/torchaudio/doc-assets/TACOTRON2_GRIFFINLIM_CHAR_LJSPEECH_v2.png :alt: Spectrogram generated by Tacotron2 .. raw:: html z3tacotron2_english_phonemes_1500_epochs_ljspeech.pth`a7Phoneme-based TTS pipeline with :py:class:`torchaudio.models.Tacotron2` and :py:class:`torchaudio.transforms.GriffinLim`. The text processor encodes the input texts based on phoneme. It uses `DeepPhonemizer `__ to convert graphemes to phonemes. The model (*en_us_cmudict_forward*) was trained on `CMUDict `__. Tacotron2 was trained on *LJSpeech* [:footcite:`ljspeech17`] for 1,500 epochs. You can find the training script `here `__. The text processor is set to the *"english_phonemes"*. The vocoder is based on :py:class:`torchaudio.transforms.GriffinLim`. Please refer to :func:`torchaudio.pipelines.Tacotron2TTSBundle` for the usage. Example - "Hello world! T T S stands for Text to Speech!" .. image:: https://download.pytorch.org/torchaudio/doc-assets/TACOTRON2_GRIFFINLIM_PHONE_LJSPEECH.png :alt: Spectrogram generated by Tacotron2 .. raw:: html Example - "The examination and testimony of the experts enabled the Commission to conclude that five shots may have been fired," .. image:: https://download.pytorch.org/torchaudio/doc-assets/TACOTRON2_GRIFFINLIM_PHONE_LJSPEECH_v2.png :alt: Spectrogram generated by Tacotron2 .. raw:: html z=tacotron2_english_characters_1500_epochs_wavernn_ljspeech.pthz%wavernn_10k_epochs_8bits_ljspeech.pth)rorpr{r|aiCharacter-based TTS pipeline with :py:class:`torchaudio.models.Tacotron2` and :py:class:`torchaudio.models.WaveRNN`. The text processor encodes the input texts character-by-character. Tacotron2 was trained on *LJSpeech* [:footcite:`ljspeech17`] for 1,500 epochs. You can find the training script `here `__. The following parameters were used; ``win_length=1100``, ``hop_length=275``, ``n_fft=2048``, ``mel_fmin=40``, and ``mel_fmax=11025``. The vocder is based on :py:class:`torchaudio.models.WaveRNN`. It was trained on 8 bits depth waveform of *LJSpeech* [:footcite:`ljspeech17`] for 10,000 epochs. You can find the training script `here `__. Please refer to :func:`torchaudio.pipelines.Tacotron2TTSBundle` for the usage. Example - "Hello world! T T S stands for Text to Speech!" .. image:: https://download.pytorch.org/torchaudio/doc-assets/TACOTRON2_WAVERNN_CHAR_LJSPEECH.png :alt: Spectrogram generated by Tacotron2 .. raw:: html Example - "The examination and testimony of the experts enabled the Commission to conclude that five shots may have been fired," .. image:: https://download.pytorch.org/torchaudio/doc-assets/TACOTRON2_WAVERNN_CHAR_LJSPEECH_v2.png :alt: Spectrogram generated by Tacotron2 .. raw:: html z;tacotron2_english_phonemes_1500_epochs_wavernn_ljspeech.ptha?Phoneme-based TTS pipeline with :py:class:`torchaudio.models.Tacotron2` and :py:class:`torchaudio.models.WaveRNN`. The text processor encodes the input texts based on phoneme. It uses `DeepPhonemizer `__ to convert graphemes to phonemes. The model (*en_us_cmudict_forward*) was trained on `CMUDict `__. Tacotron2 was trained on *LJSpeech* [:footcite:`ljspeech17`] for 1,500 epochs. You can find the training script `here `__. The following parameters were used; ``win_length=1100``, ``hop_length=275``, ``n_fft=2048``, ``mel_fmin=40``, and ``mel_fmax=11025``. The vocder is based on :py:class:`torchaudio.models.WaveRNN`. It was trained on 8 bits depth waveform of *LJSpeech* [:footcite:`ljspeech17`] for 10,000 epochs. You can find the training script `here `__. Please refer to :func:`torchaudio.pipelines.Tacotron2TTSBundle` for the usage. Example - "Hello world! T T S stands for Text to Speech!" .. image:: https://download.pytorch.org/torchaudio/doc-assets/TACOTRON2_WAVERNN_PHONE_LJSPEECH.png :alt: Spectrogram generated by Tacotron2 .. raw:: html Example - "The examination and testimony of the experts enabled the Commission to conclude that five shots may have been fired," .. image:: https://download.pytorch.org/torchaudio/doc-assets/TACOTRON2_WAVERNN_PHONE_LJSPEECH_v2.png :alt: Spectrogram generated by Tacotron2 .. raw:: html )4rCZ dataclassesrtypingrrrrrrrVr Ztorchaudio._internalr Ztorchaudio.functionalr Ztorchaudio.modelsr r Ztorchaudio.transformsrrrBrZ interfacer__all__rsrkrr9nnModuleZVocoderrIr^rgrlrnrzrrrrrZ_get_taco_paramsZ"TACOTRON2_GRIFFINLIM_CHAR_LJSPEECH__doc__Z#TACOTRON2_GRIFFINLIM_PHONE_LJSPEECHZ_get_wrnn_paramsZTACOTRON2_WAVERNN_CHAR_LJSPEECHZ TACOTRON2_WAVERNN_PHONE_LJSPEECHrrrrsn       &   & + )