text-to-speech-en-0001 (composite)

Use Case and High-Level Description

This is a speech synthesis composite model that simultaneously reconstructs mel-spectrogram and wave form from text. The model generates wave form from symbol sequences separated by space. The model is built on top of the modified ForwardTacotron and modified MelGAN frameworks.

Composite model specification

Metric Value
Source framework PyTorch*

Duration prediction model specification

The text-to-speech-en-0001-duration-prediction model is a ForwardTacotron-based duration predictor for symbols.

Metric Value
GFlops 15.84
MParams 13.569

Inputs

Sequence, name: input_seq, shape: 1, 512, format: B,C where:

  • B - batch size
  • C - number of symbols in sequence

Sequence, name: input_mask, shape: 1, 1, 512, format: B, D, C where:

  • B - batch size
  • D - extra dimension for multiplication
  • C - number of symbols in sequence

Mask for input sequence, name: input_mask, shape: 1, 1, 512, format: B, D, C where:

  • B - batch size
  • D - extra dimension for multiplication
  • C - number of symbols in sequence

Mask for relative position representation in attention, name: pos_mask, shape: 1, 1, 512, 512, format: B, D, C, C where:

  • B - batch size
  • D - extra dimension for multiplication
  • C - number of symbols in sequence

Outputs

  1. Duration for input symbols, name: duration, shape: 1, 512, 1, format B, C, H. Contains predicted duration for each of the symbol in sequence.
    • B - batch size
    • C - number of symbols in sequence
    • H - empty dimension
  2. Processed embeddings, name: embeddings, shape: 1, 512, 256, format BxCxH. Contains processed embeddings for each symbol in sequence.
    • B - batch size
    • C - number of symbols in sequence
    • H - height of the intermediate feature map

Mel-spectrogram regression model specification

The text-to-speech-en-0001-regression model accepts aligned by duration processed embeddings (for example: if duration is [2, 3] and processed embeddings is [[1, 2], [3, 4]], aligned embeddings is [[1, 2], [1, 2], [1,2], [3, 4], [3, 4]]) and produces mel-spectrogram.

Metric Value
GFlops 7.65
MParams 4.96

Inputs

Processed embeddigs aligned by durations, name: data, shape: 1x512x256, format: BxTxC where:

  • B - batch size
  • T - time in mel-spectrogram
  • C - processed embedding dimension

Mask for 'data' by time dimension, name: data_mask, shape: 1x1x512, format: BxDxT where:

  • B - batch size
  • D - extra dimension for multiplication
  • T - time in mel-spectrogram

Mask for relative position representation in attention, name: pos_mask, shape: 1x1x512x512, format: BxDxCxC where:

  • B - batch size
  • D - extra dimension for multiplication
  • C - number of symbols in sequence

Output

Mel-spectrogram, name: mel, shape: 80x512, format: CxT where:

  • T - time in mel-spectrogram
  • C - number of rows in mel-spectrogram

Audio generation model specification

The text-to-speech-en-0001-generation model is a MelGAN based audio generator.

Metric Value
GFlops 48.38
MParams 12.77

Inputs

Mel-spectrogram, name: mel, shape: 1x80x128, format: BxCxT where:

  • B - batch size
  • C - number of rows in mel-spectrogram
  • T - time in mel-spectrogram

Outputs

Audio, name: audio, shape: 32768, format: T where:

  • T - time in audio with sampling rate 22050 (~1.5 sec).

Legal Information

[*] Other names and brands may be claimed as the property of others.