Skip to content

Model_Architectures.md

Latest commit

 

History

History
74 lines (43 loc) · 2.66 KB

Model_Architectures.md

File metadata and controls

74 lines (43 loc) · 2.66 KB
graph LR
    BaseModel["BaseModel"]
    Model_Configuration["Model Configuration"]
    Acoustic_Models["Acoustic Models"]
    Vocoder_Models["Vocoder Models"]
    Inference_Wrappers["Inference Wrappers"]
    Model_Configuration -- "instantiates" --> Acoustic_Models
    Model_Configuration -- "instantiates" --> Vocoder_Models
    Acoustic_Models -- "inherits from" --> BaseModel
    Vocoder_Models -- "inherits from" --> BaseModel
    Acoustic_Models -- "provides spectrogram to" --> Vocoder_Models
    Inference_Wrappers -- "wraps for deployment" --> Acoustic_Models
    Inference_Wrappers -- "wraps for deployment" --> Vocoder_Models
Loading

CodeBoardingDemoContact

Details

An analysis of the Model Architectures subsystem, revealing a modular and extensible design centered around a core base class and a clear separation between acoustic and vocoder models.

BaseModel

An abstract base class that defines the common interface for all TTS models. It establishes the contract for the forward pass, loss calculation, and inference, enabling any model to be seamlessly integrated with the Trainer Engine.

Related Classes/Methods:

  • tensorflow_tts.models.base_model

Model Configuration

A set of data classes used to define the hyperparameters and architecture of each model. These configuration objects are used to instantiate models, promoting a clean separation of configuration from code.

Related Classes/Methods:

  • tensorflow_tts.configs

Acoustic Models

A category of models responsible for the first stage of synthesis: converting processed text inputs into a mel-spectrogram representation.

Related Classes/Methods:

  • fastspeech2
  • tacotron2
  • fastspeech

Vocoder Models

A category of models that perform the second stage of synthesis: converting mel-spectrograms into an audible waveform. These often use GANs, comprising both a generator and a discriminator.

Related Classes/Methods:

  • melgan
  • mb_melgan
  • hifigan

Inference Wrappers

Specialized, serializable tf.Module classes that wrap trained models to provide a simplified, consistent inference interface (tf.function decorated __call__).

Related Classes/Methods:

  • SavableTFFastSpeech

FAQ