forward-tacotron (composite)

Use Case and High-Level Description

ForwardTacotron is a model for the text-to-speech task originally trained in PyTorch* then converted to ONNX* format. The model was trained on LJSpeech dataset. ForwardTacotron performs mel-spectrogram regression from text. For details see paper, paper, repository.

ONNX Models

We provide pre-trained models in ONNX format for user convenience.

Steps to Reproduce training in PyTorch and Conversion to ONNX

Model is provided in ONNX format, which was obtained by the following steps.

  1. Clone the original repository

    git clone https://github.com/as-ideas/ForwardTacotron
    cd ForwardTacotron
  2. Checkout the commit that the conversion was tested on:

    git checkout 78789c1aa845057bb2f799e702b1be76bf7defd0
  3. Follow README.md and train ForwardTacotron model.

  4. Copy provided script forward_to_onnx.py to ForwardTacotron root directory.

  5. Run provided script for conversion ForwardTacotron to onnx format

    python3 forward_to_onnx.py --tts_weights checkpoints/ljspeech_tts.forward/fast_speech_step<iteration>K_weights.pyt

    Notes:

    1. Since ONNX doesn’t support the build_index operation from PyTorch pipeline, the model is divided into two parts: forward_tacotron_duration_prediction.onnx, forward_tacotron_regression.onnx.

    2. We stopped training of the Tacotron model in 183K iteration for alignment generation and stopped ForwardTacotron training in 290K iteration.

Composite model specification

Metric

Value

Source framework

PyTorch*

Accuracy

Subjective

forward-tacotron-duration-prediction model specification

The forward-tacotron-duration-prediction model accepts preprocessed text (see text_to_sequence in repository) and produces processed embeddings and duration in time for every processed embedding.

Metric

Value

GOPs

6.66

MParams

13.81

Input

Sequence, name: input_seq, shape: 1, 241, format: B, C, where:

  • B - batch size

  • C - number of symbols in sequence (letters or phonemes)

Output

  1. Duration for input symbols, name: duration, shape: 1, 241, 1, format B, C, H. Contains predicted duration for each of the symbol in sequence.

    • B - batch size

    • C - number of symbols in sequence (letters or phonemes)

    • H - empty dimension

  2. Processed embeddings, name: embeddings, shape: 1, 241, 512, format B, C, H. Contains processed embeddings for each symbol in sequence.

    • B - batch size

    • C - number of symbols in sequence (letters or phonemes)

    • H - height of the intermediate feature map

forward-tacotron-regression model specification

The forward-tacotron-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

GOPs

4.91

MParams

3.05

Input

Processed embeddings aligned by durations, name: data, shape: 1, 805, 512, format: B, T, C, where:

  • B - batch size

  • T - time in mel-spectrogram

  • C - processed embedding dimension

Output

Mel-spectrogram, name: mel, shape: 80, 805, format: C, T, where:

  • T - time in mel-spectrogram

  • C - number of mels in mel-spectrogram

Download a Model and Convert it into OpenVINO™ IR Format

You can download models and if necessary convert them into OpenVINO™ IR format using the Model Downloader and other automation tools as shown in the examples below.

An example of using the Model Downloader:

omz_downloader --name <model_name>

An example of using the Model Converter:

omz_converter --name <model_name>

Demo usage

The model can be used in the following demos provided by the Open Model Zoo to show its capabilities: