FastPitch 1.1 for PyTorch

This repository provides a script and recipe to train the FastPitch model to achieve state-of-the-art accuracy and is tested and maintained by NVIDIA.

Table Of Contents

Model overview
Setup
- Requirements
Quick Start Guide
Advanced
Performance
- Benchmarking
  - Training performance benchmark
  - Inference performance benchmark
- Results
Release notes
- Changelog
- Known issues

Model overview

FastPitch is one of two major components in a neural, text-to-speech (TTS) system:

a mel-spectrogram generator such as FastPitch or Tacotron 2, and
a waveform synthesizer such as WaveGlow (refer to NVIDIA example code).

Such a two-component TTS system is able to synthesize natural-sounding speech from raw transcripts.

The FastPitch model generates mel-spectrograms and predicts a pitch contour from raw input text. In version 1.1, it does not need any pre-trained aligning model to bootstrap from. It allows exerting additional control over the synthesized utterances, such as:

modify the pitch contour to control the prosody,
increase or decrease the fundamental frequency in a natural sounding way, that preserves the perceived identity of the speaker,
alter the rate of speech,
adjust the energy,
specify input as graphemes or phonemes,
switch speakers when the model has been trained with data from multiple speakers. Some of the capabilities of FastPitch are presented on the website with samples.

Speech synthesized with FastPitch has state-of-the-art quality, and does not suffer from missing/repeating phrases as Tacotron 2 does. This is reflected in Mean Opinion Scores (details).

Model	Mean Opinion Score (MOS)
Tacotron 2	3.946 ± 0.134
FastPitch 1.0	4.080 ± 0.133

The current version of the model offers even higher quality, as reflected in the pairwise preference scores (details).

Model	Average preference
FastPitch 1.0	0.435 ± 0.068
FastPitch 1.1	0.565 ± 0.068

The FastPitch model is based on the FastSpeech model. The main differences between FastPitch and FastSpeech are that FastPitch:

no dependence on external aligner (Transformer TTS, Tacotron 2); in version 1.1, FastPitch aligns audio to transcriptions by itself as in One TTS Alignment To Rule Them All,
explicitly learns to predict the pitch contour,
pitch conditioning removes harsh sounding artifacts and provides faster convergence,
no need for distilling mel-spectrograms with a teacher model,
capabilities to train a multi-speaker model.

The FastPitch model is similar to FastSpeech2, which has been developed concurrently. FastPitch averages pitch/energy values over input tokens, and treats energy as optional.

FastPitch is trained on a publicly available LJ Speech dataset.

This model is trained with mixed precision using Tensor Cores on NVIDIA Volta, NVIDIA Turing, and the NVIDIA Ampere GPU architectures. Therefore, researchers can get results from 2.0x to 2.7x faster than training without Tensor Cores while experiencing the benefits of mixed precision training. This model is tested against each NGC monthly container release to ensure consistent accuracy and performance over time.

Model architecture

FastPitch is a fully feedforward Transformer model that predicts mel-spectrograms from raw text (Figure 1). The entire process is parallel, which means that all input letters are processed simultaneously to produce a full mel-spectrogram in a single forward pass.

Figure 1. Architecture of FastPitch (source). The model is composed of a bidirectional Transformer backbone (also known as a Transformer encoder), a pitch predictor, and a duration predictor. After passing through the first *N* Transformer blocks, encoding, the signal is augmented with pitch information and discretely upsampled. Then it goes through another set of *N* Transformer blocks, with the goal of smoothing out the upsampled signal and constructing a mel-spectrogram.

Default configuration

The FastPitch model supports multi-GPU and mixed precision training with dynamic loss scaling (refer to Apex code here), as well as mixed precision inference.

The following features were implemented in this model:

data-parallel multi-GPU training,
dynamic loss scaling with backoff for Tensor Cores (mixed precision) training,
gradient accumulation for reproducible results regardless of the number of GPUs.

Pitch contours and mel-spectrograms can be generated online during training. To speed-up training, those could be generated during the pre-processing step and read directly from the disk during training. For more information on data pre-processing, refer to Dataset guidelines and the paper.

Feature support matrix

The following features are supported by this model.

Feature	FastPitch
Automatic mixed precision (AMP)	Yes
Distributed data parallel (DDP)	Yes

Features

Automatic Mixed Precision (AMP) - This implementation uses native PyTorch AMP implementation of mixed precision training. It allows us to use FP16 training with FP32 master weights by modifying just a few lines of code.

DistributedDataParallel (DDP) - The model uses PyTorch Lightning implementation of distributed data parallelism at the module level, which can run across multiple machines.

Mixed precision training

Mixed precision is the combined use of different numerical precisions in a computational method. Mixed precision training offers significant computational speedup by performing operations in half-precision format while storing minimal information in single-precision to retain as much information as possible in critical parts of the network. Since the introduction of Tensor Cores in NVIDIA Volta, and following with both the Turing and Ampere architectures, significant training speedups are experienced by switching to mixed precision -- up to 3x overall speedup on the most arithmetically intense model architectures. Using mixed precision training requires two steps:

Porting the model to use the FP16 data type where appropriate.
Adding loss scaling to preserve small gradient values.

The ability to train deep learning networks with lower precision was introduced in the Pascal architecture and first supported in CUDA 8 in the NVIDIA Deep Learning SDK.

For information about:

How to train using mixed precision, refer to the Mixed Precision Training paper and Training With Mixed Precision documentation.
Techniques used for mixed precision training, refer to the Mixed-Precision Training of Deep Neural Networks blog.
APEX tools for mixed precision training, refer to the NVIDIA Apex: Tools for Easy Mixed-Precision Training in PyTorch.

Enabling mixed precision

For training and inference, mixed precision can be enabled by adding the --amp flag. Mixed precision is using native PyTorch implementation.

Enabling TF32

TensorFloat-32 (TF32) is the new math mode in NVIDIA A100 GPUs for handling the matrix math, also called tensor operations. TF32 running on Tensor Cores in A100 GPUs can provide up to 10x speedups compared to single-precision floating-point math (FP32) on Volta GPUs.

TF32 Tensor Cores can speed up networks using FP32, typically with no loss of accuracy. It is more robust than FP16 for models which require a high dynamic range for weights or activations.

For more information, refer to the TensorFloat-32 in the A100 GPU Accelerates AI Training, HPC up to 20x blog post.

TF32 is supported in the NVIDIA Ampere GPU architecture and is enabled by default.

Glossary

Character duration The time during which a character is being articulated. It could be measured in milliseconds, mel-spectrogram frames, and so on. Some characters are not pronounced, and thus, have 0 duration.

Fundamental frequency The lowest vibration frequency of a periodic soundwave, for example, is produced by a vibrating instrument, and it is perceived as the loudest. In the context of speech, it refers to the frequency of vibration of vocal cords. It is abbreviated as f0.

Pitch A perceived frequency of vibration of music or sound.

Transformer The paper Attention Is All You Need introduces a novel architecture called Transformer, which repeatedly applies the attention mechanism. It transforms one sequence into another.

Setup

The following section lists the requirements that you need to meet in order to start training the FastPitch model.

Requirements

This repository contains Dockerfile that extends the PyTorch NGC container and encapsulates some dependencies. Aside from these dependencies, ensure you have the following components:

NVIDIA Docker
PyTorch 22.08-py3 NGC container or newer
supported GPUs:

For more information about how to get started with NGC containers, refer to the following sections from the NVIDIA GPU Cloud Documentation and the Deep Learning Documentation:

For those unable to use the PyTorch NGC container, to set up the required environment or create your own container, refer to the versioned NVIDIA Container Support Matrix.

Quick Start Guide

To train your model using mixed or TF32 precision with Tensor Cores or using FP32, perform the following steps using the default parameters of the FastPitch model on the LJSpeech 1.1 dataset. For the specifics concerning training and inference, refer to the Advanced section. Pre-trained FastPitch models are available for download on NGC.

Clone the repository.

git clone https://github.com/NVIDIA/DeepLearningExamples.git
cd DeepLearningExamples/PyTorch/SpeechSynthesis/FastPitch

Build and run the FastPitch PyTorch NGC container.

By default, the container will use all available GPUs.
```
bash scripts/docker/build.sh
bash scripts/docker/interactive.sh
```

Download and preprocess the dataset.

Use the scripts to automatically download and preprocess the training, validation, and test datasets:

bash scripts/download_dataset.sh
bash scripts/prepare_dataset.sh

The data is downloaded to the ./LJSpeech-1.1 directory (on the host). The ./LJSpeech-1.1 directory is mounted under the /workspace/fastpitch/LJSpeech-1.1 location in the NGC container. The complete dataset has the following structure:

./LJSpeech-1.1
├── mels             # (optional) Pre-calculated target mel-spectrograms; can be calculated online
├── metadata.csv     # Mapping of waveforms to utterances
├── pitch            # Fundamental frequency contours for input utterances; can be calculated online
├── README
└── wavs             # Raw waveforms

To run inference in mixed precision, use the --amp flag. The output audio will be stored in the path specified by the -o argument. Consult the inference.py to learn more options, such as setting the batch size.

Advanced

The following sections provide greater details of the dataset, running training and inference, and the training results.

Scripts and sample code

The repository holds code for FastPitch (training and inference) and WaveGlow (inference only). The code specific to a particular model is located in that model’s directory - ./fastpitch and ./waveglow - and common functions live in the ./common directory. The model-specific scripts are as follows:

<model_name>/model.py - the model architecture, definition of forward and inference functions
<model_name>/arg_parser.py - argument parser for parameters specific to a given model
<model_name>/data_function.py - data loading functions
<model_name>/loss_function.py - loss function for the model

In the root directory ./ of this repository, the ./train.py script is used for training, while inference can be executed with the ./inference.py script. The script ./models.py is used to construct a model of the requested type and properties.

The repository is structured similarly to the NVIDIA Tacotron2 Deep Learning example so that they could be combined in more advanced use cases.

Parameters

In this section, we list the most important hyperparameters and command-line arguments, together with their default values that are used to train FastPitch.

--epochs - number of epochs (default: 1000)
--learning-rate - learning rate (default: 0.1)
--batch-size - batch size for a single forward-backward step (default: 16)
--grad-accumulation - number of steps over which gradients are accumulated (default: 2)
--amp - use mixed precision training (default: disabled)
--load-pitch-from-disk - pre-calculated fundamental frequency values, estimated before training, are loaded from the disk during training (default: enabled)
--energy-conditioning - enables additional conditioning on energy (default: enabled)
--p-arpabet - probability of choosing phonemic over graphemic representation for every word, if available (default: 1.0)

Command-line options

To review the full list of available options and their descriptions, use the -h or --help command-line option, for example:

python train.py --help

The following example output is printed when running the model:

DLL 2021-06-14 23:08:53.659718 - epoch    1 | iter   1/48 | loss 40.97 | mel loss 35.04 | kl loss 0.02240 | kl weight 0.01000 |    5730.98 frames/s | took 24.54 s | lrate 3.16e-06
DLL 2021-06-14 23:09:28.449961 - epoch    1 | iter   2/48 | loss 41.07 | mel loss 35.12 | kl loss 0.02258 | kl weight 0.01000 |    4154.18 frames/s | took 34.79 s | lrate 6.32e-06
DLL 2021-06-14 23:09:59.365398 - epoch    1 | iter   3/48 | loss 40.86 | mel loss 34.93 | kl loss 0.02252 | kl weight 0.01000 |    4589.15 frames/s | took 30.91 s | lrate 9.49e-06

Getting the data

The FastPitch and WaveGlow models were trained on the LJSpeech-1.1 dataset. The ./scripts/download_dataset.sh script will automatically download and extract the dataset to the ./LJSpeech-1.1 directory.

Dataset guidelines

The LJSpeech dataset has 13,100 clips that amount to about 24 hours of speech of a single female speaker. Since the original dataset does not define a train/dev/test split of the data, we provide a split in the form of three file lists:

./filelists
├── ljs_audio_pitch_text_train_v3.txt
├── ljs_audio_pitch_text_test.txt
└── ljs_audio_pitch_text_val.txt

FastPitch predicts character durations just as FastSpeech does. FastPitch 1.1 aligns input symbols to output mel-spectrogram frames automatically and does not rely on any external aligning model. FastPitch training can now be started on raw waveforms without any pre-processing: pitch values and mel-spectrograms will be calculated online.

For every mel-spectrogram frame, its fundamental frequency in Hz is estimated with the Probabilistic YIN algorithm.

Figure 2. Pitch estimates for mel-spectrogram frames of the phrase "in being comparatively" (in blue) averaged over characters (in red). Silent letters have a duration of 0 and are omitted.

Multi-dataset

Follow these steps to use datasets different from the default LJSpeech dataset.

Prepare a directory with .wav files.
```
./my_dataset
└── wavs
```
Prepare filelists with transcripts and paths to .wav files. They define the training/validation split of the data (the test is currently unused):
```
./filelists
├── my-dataset_audio_text_train.txt
└── my-dataset_audio_text_val.txt
```
Those filelists should list a single utterance per line as:
```
`<audio file path>|<transcript>`
```
The <audio file path> is the relative path to the path provided by the --dataset-path option of train.py.

Run the pre-processing script to calculate pitch:

 python prepare_dataset.py \
     --wav-text-filelists filelists/my-dataset_audio_text_train.txt \
                          filelists/my-dataset_audio_text_val.txt \
     --n-workers 16 \
     --batch-size 1 \
     --dataset-path $DATA_DIR \
     --extract-pitch \
     --f0-method pyin

Prepare file lists with paths to pre-calculated pitch:

./filelists
├── my-dataset_audio_pitch_text_train.txt
└── my-dataset_audio_pitch_text_val.txt

In order to use the prepared dataset, pass the following to the train.py script:

--dataset-path ./my_dataset` \
--training-files ./filelists/my-dataset_audio_pitch_text_train.txt \
--validation files ./filelists/my-dataset_audio_pitch_text_val.txt

Training process

FastPitch is trained to generate mel-spectrograms from raw text input. It uses short-time Fourier transform (STFT) to generate target mel-spectrograms from audio waveforms to be the training targets.

The training loss is averaged over an entire training epoch, whereas the validation loss is averaged over the validation dataset. Performance is reported in total output mel-spectrogram frames per second and recorded as train_frames/s (after each iteration) and avg_train_frames/s (averaged over epoch) in the output log file ./output/nvlog.json. The result is averaged over an entire training epoch and summed over all GPUs that were included in the training.

The scripts/train.sh script is configured for 8x GPU with at least 16GB of memory:

--batch-size 16
--grad-accumulation 2

In a single accumulated step, there are batch_size x grad_accumulation x GPUs = 256 examples being processed in parallel. With a smaller number of GPUs, increase --grad_accumulation to keep this relation satisfied, e.g., through env variables

NUM_GPUS=1 GRAD_ACCUMULATION=16 bash scripts/train.sh

Inference process

You can run inference using the ./inference.py script. This script takes text as input and runs FastPitch and then WaveGlow inference to produce an audio file. It requires pre-trained checkpoints of both models and input text as a text file, with one phrase per line.

Pre-trained FastPitch models are available for download on NGC.

Having pre-trained models in place, run the sample inference on LJSpeech-1.1 test-set with:

bash scripts/inference_example.sh

Examine the inference_example.sh script to adjust paths to pre-trained models, and call python inference.py --help to learn all available options. By default, synthesized audio samples are saved in ./output/audio_* folders.

FastPitch allows us to linearly adjust the rate of synthesized speech like FastSpeech. For instance, pass --pace 0.5 for a twofold decrease in speed.

For every input character, the model predicts a pitch cue - an average pitch over a character in Hz. Pitch can be adjusted by transforming those pitch cues. A few simple examples are provided below.

Transformation	Flag	Samples
-	-	link
Amplify pitch wrt. to the mean pitch	`--pitch-transform-amplify`	link
Invert pitch wrt. to the mean pitch	`--pitch-transform-invert`	link
Raise/lower pitch by	`--pitch-transform-shift <hz>`	link
Flatten the pitch to a constant value	`--pitch-transform-flatten`	link
Change the rate of speech (1.0 = unchanged)	`--pace <value>`	link

The flags can be combined. Modify these functions directly in the inference.py script to gain more control over the final result.

You can find all the available options by callng python inference.py --help. More examples are presented on the website with samples.

Example: Training a model on Mandarin Chinese

FastPitch can easily be trained or fine-tuned on datasets in various languages. We present an example of training on the Mandarin Chinese dataset capable of pronouncing phrases in English (for example, brand names). For an overview of the deployment of this model in Chunghwa Telecom, refer to the blogpost (in Chinese).

Set up the repository and run a Docker container

Follow stetps 1. and 2. of the Quick Start Guide.
Download the data

The dataset for this section has been provided by Chunghwa Telecom Laboratories and is available for download on NGC under the CC BY-NC 4.0 license.

The dataset can be downloaded manually after signing in to NGC as files.zip or SF_bilingual.zip, depending on the method (manual or via command line). Afterward, it has to be pre-processed to extract pitch for training and prepare train/dev/test filelists:
```
pip install -r scripts/mandarin_chinese/requirements.txt
bash scripts/mandarin_chinese/prepare_dataset.sh path/to/files.zip
```
The procedure should take about half an hour. If it completes successfully, ./data/SF_bilingual prepared successfully. will be written to the standard output.

After pre-processing, the dataset will be located at ./data/SF_bilingual, and training/inference filelists at ./filelists/sf_*.

Add support for textual inputs in the target language.

The model is trained end-to-end, and supporting a new language requires to specify the input symbol set, text normalization routines, and (optionally) grapheme-to-phoneme (G2P) conversion for phoneme-based synthesis. Our main modifications touch the following files:

./common/text
├── symbols.py
├── text_processing.py
└── zh
    ├── chinese.py
    ├── mandarin_text_processing.py
    └── pinyin_dict.txt

We make small changes to symbols.py and text_processing.py and keep the crucial code in the zh directory.

We design our Mandarin Chinese symbol set as an extension of the English symbol set, appending to symbols lists of _mandarin_phonemes and _chinese_punctuation:

# common/text/symbols.py

def get_symbols(symbol_set='english_basic'):

    # ...

    elif symbol_set == 'english_mandarin_basic':
        from .zh.chinese import chinese_punctuations, valid_symbols as mandarin_valid_symbols

        # Prepend "#" to mandarin phonemes to ensure uniqueness (some are the same as uppercase letters):
        _mandarin_phonemes = ['#' + s for s in mandarin_valid_symbols]

        _pad = '_'
        _punctuation = '!\'(),.:;? '
        _chinese_punctuation = ["#" + p for p in chinese_punctuations]
        _special = '-'
        _letters = 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz'
        symbols = list(_pad + _special + _punctuation + _letters) + _arpabet + _mandarin_phonemes + _chinese_punctuation

Text normalization and G2P are performed by a TextProcessing instance. We implement Mandarin text processing inside a MandarinTextProcessing class. For G2P, an off-shelf pypinyin phonemizer and the CMU Dictionary are used. MandarinTextProcessing is applied to the data only if english_mandarin_basic symbol set is in use:

# common/text/text_processing.py

def get_text_processing(symbol_set, text_cleaners, p_arpabet):
    if symbol_set in ['englh_basic', 'english_basic_lowercase', 'english_expanded']:
        return TextProcessing(symbol_set, text_cleaners, p_arpabet=p_arpabet)
    elif symbol_set == 'english_mandarin_basic':
        from common.text.zh.mandarin_text_processing import MandarinTextProcessing
        return MandarinTextProcessing(symbol_set, text_cleaners, p_arpabet=p_arpabet)

Note that text normalization is dependent on the target language, domain, and assumptions on how normalized the input already is.

Train the model

The SF dataset is rather small (4.5 h compared to 24 h in LJSpeech-1.1). There are numerous English phrases in the transcriptions, such as technical terms and proper nouns. Thus, it is beneficial to initialize model weights with a pre-trained English model from NGC, using the flag --init-from-checkpoint.

Note that by initializing with another model, possibly trained on a different symbol set, we also initialize grapheme/phoneme embedding tables. For this reason, we design the english_mandarin_basic symbol set as an extension of english_basic, so that the same English phonemes would retain their embeddings.

In order to train, issue
```
NUM_GPUS=<available_gpus> GRAD_ACCUMULATION=<number> bash scripts/mandarin_chinese/train.sh
```
Adjust the variables to satisfy $NUM_GPUS x $GRAD_ACCUMULATION = 256.

The model will be trained for 1000 epochs. Note that we have disabled mixed-precision training, as we found it unstable at times on this dataset.
Synthesize

After training, samples can be synthesized (audio sample):
```
bash scripts/mandarin_chinese/inference.sh
```
Paths to specific checkpoints can be supplied as env variables or changed directly in the .sh files.

Performance

Benchmarking

The following section shows how to run benchmarks measuring the model performance in training and inference mode.

Training performance benchmark

To benchmark the training performance on a specific batch size, run:

NVIDIA DGX A100 (8x A100 80GB)

    AMP=true NUM_GPUS=1 BS=32 GRAD_ACCUMULATION=8 EPOCHS=10 bash scripts/train.sh
    AMP=true NUM_GPUS=8 BS=32 GRAD_ACCUMULATION=1 EPOCHS=10 bash scripts/train.sh
    AMP=false NUM_GPUS=1 BS=32 GRAD_ACCUMULATION=8 EPOCHS=10 bash scripts/train.sh
    AMP=false NUM_GPUS=8 BS=32 GRAD_ACCUMULATION=1 EPOCHS=10 bash scripts/train.sh

NVIDIA DGX-1 (8x V100 16GB)

    AMP=true NUM_GPUS=1 BS=16 GRAD_ACCUMULATION=16 EPOCHS=10 bash scripts/train.sh
    AMP=true NUM_GPUS=8 BS=16 GRAD_ACCUMULATION=2 EPOCHS=10 bash scripts/train.sh
    AMP=false NUM_GPUS=1 BS=16 GRAD_ACCUMULATION=16 EPOCHS=10 bash scripts/train.sh
    AMP=false NUM_GPUS=8 BS=16 GRAD_ACCUMULATION=2 EPOCHS=10 bash scripts/train.sh

Each of these scripts runs for 10 epochs, and for each epoch, measures the average number of items per second. The performance results can be read from the nvlog.json files produced by the commands.

Inference performance benchmark

To benchmark the inference performance on a specific batch size, run:

For FP16

AMP=true BS_SEQUENCE=”1 4 8” REPEATS=100 bash scripts/inference_benchmark.sh

For FP32 or TF32

AMP=false BS_SEQUENCE=”1 4 8” REPEATS=100 bash scripts/inference_benchmark.sh

The output log files will contain performance numbers for the FastPitch model (number of output mel-spectrogram frames per second, reported as generator_frames/s w ) and for WaveGlow (nuber of output samples per second, reported as waveglow_samples/s). The inference.py script will run a few warm-up iterations before running the benchmark. Inference will be averaged over 100 runs, as set by the REPEATS env variable.

Results

The following sections provide details on how we achieved our performance and accuracy in training and inference.

Training accuracy results

Training accuracy: NVIDIA DGX A100 (8x A100 80GB)

Our results were obtained by running the ./platform/DGXA100_FastPitch_{AMP,TF32}_8GPU.sh training script in the PyTorch 21.05-py3 NGC container on NVIDIA DGX A100 (8x A100 80GB) GPUs.

Loss (Model/Epoch)	50	250	500	750	1000	1250	1500
FastPitch AMP	3.35	2.89	2.79	2.71	2.68	2.64	2.61
FastPitch TF32	3.37	2.88	2.78	2.71	2.68	2.63	2.61

Training accuracy: NVIDIA DGX-1 (8x V100 16GB)

Our results were obtained by running the ./platform/DGX1_FastPitch_{AMP,FP32}_8GPU.sh training script in the PyTorch 21.05-py3 NGC container on NVIDIA DGX-1 with 8x V100 16GB GPUs.

All of the results were produced using the train.py script as described in the Training process section of this document.

Loss (Model/Epoch)	50	250	500	750	1000	1250	1500
FastPitch AMP	3.38	2.88	2.79	2.71	2.68	2.64	2.61
FastPitch FP32	3.38	2.89	2.80	2.71	2.68	2.65	2.62

Training performance results

Training performance: NVIDIA DGX A100 (8x A100 80GB)

Our results were obtained by running the ./platform/DGXA100_FastPitch_{AMP,TF32}_8GPU.sh training script in the PyTorch 22.08-py3 NGC container on NVIDIA DGX A100 (8x A100 80GB) GPUs. Performance numbers, in output mel-scale spectrogram frames per second, were averaged over an entire training epoch.

Batch size / GPU	GPUs	Grad accumulation	Throughput - TF32	Throughput - mixed precision	Throughput speedup (TF32 to mixed precision)	Strong scaling - TF32	Strong scaling - mixed precision
128	1	2	141,028	148,149	1.05	1.00	1.00
64	4	1	525,879	614,857	1.17	3.73	4.15
32	8	1	914,350	1,022,722	1.12	6.48	6.90

Expected training time

The following table shows the expected training time for convergence for 1000 epochs:

Batch size / GPU	GPUs	Grad accumulation	Time to train with TF32 (Hrs)	Time to train with mixed precision (Hrs)	Speed-up with mixed precision
128	1	2	14.5	13.8	1.05
64	4	1	4.1	3.3	1.17
32	8	1	2.2	2.0	1.12

Training performance: NVIDIA DGX-1 (8x V100 16GB)

Our results were obtained by running the ./platform/DGX1_FastPitch_{AMP,FP32}_8GPU.sh training script in the PyTorch 22.08-py3 NGC container on NVIDIA DGX-1 with 8x V100 16GB GPUs. Performance numbers, in output mel-scale spectrogram frames per second, were averaged over an entire training epoch.

Batch size / GPU	GPUs	Grad accumulation	Throughput - FP32	Throughput - mixed precision	Throughput speedup (FP32 to mixed precision)	Strong scaling - FP32	Strong scaling - mixed precision
16	1	16	31,863	83,761	2.63	1.00	1.00
16	4	4	117,971	269,143	2.28	3.70	3.21
16	8	2	225,826	435,799	1.93	7.09	5.20

To achieve these same results, follow the steps in the Quick Start Guide.

Expected training time

The following table shows the expected training time for convergence for 1000 epochs:

Batch size / GPU	GPUs	Grad accumulation	Time to train with FP32 (Hrs)	Time to train with mixed precision (Hrs)	Speed-up with mixed precision
16	1	16	64.2	24.4	2.63
16	4	4	17.4	7.6	2.28
16	8	2	9.1	4.7	1.93

Note that most of the quality is achieved after the initial 1000 epochs.

Inference performance results

The following tables show inference statistics for the FastPitch and WaveGlow text-to-speech system, gathered from 100 inference runs. Latency is measured from the start of FastPitch inference to the end of WaveGlow inference. Throughput is measured as the number of generated audio samples per second at 22KHz. RTF is the real-time factor that denotes the number of seconds of speech generated in a second of wall-clock time per input utterance. The used WaveGlow model is a 256-channel model.

Note that performance numbers are related to the length of input. The numbers reported below were taken with a moderate length of 128 characters. Longer utterances yield higher RTF, as the generator is fully parallel.

Inference performance: NVIDIA DGX A100 (1x A100 80GB)

Our results were obtained by running the ./scripts/inference_benchmark.sh inferencing benchmarking script in the PyTorch 22.08-py3 NGC container on NVIDIA DGX A100 (1x A100 80GB) GPU.

FastPitch (TorchScript, denoising)

Batch size	Precision	Avg latency (s)	Latency tolerance interval 90% (s)	Latency tolerance interval 95% (s)	Latency tolerance interval 99% (s)	Throughput (frames/sec)	Speed-up with mixed precision	Avg RTF
1	FP16	0.005	0.006	0.006	0.006	120,333	0.97	1397.07
4	FP16	0.006	0.006	0.006	0.006	424,053	1.12	1230.81
8	FP16	0.008	0.010	0.010	0.011	669,549	1.12	971.68
1	TF32	0.005	0.006	0.006	0.007	123,718	-	1436.37
4	TF32	0.007	0.007	0.007	0.007	379,980	-	1102.89
8	TF32	0.009	0.009	0.009	0.009	600,435	-	871.38

FastPitch + HiFi-GAN (TorchScript, denoising)

Batch size	Precision	Avg latency (s)	Latency tolerance interval 90% (s)	Latency tolerance interval 95% (s)	Latency tolerance interval 99% (s)	Throughput (samples/sec)	Speed-up with mixed precision	Avg RTF
1	FP16	0.015	0.016	0.016	0.016	11,431,335	1.28	518.43
4	FP16	0.038	0.040	0.040	0.040	17,670,528	1.42	200.35
8	FP16	0.069	0.069	0.070	0.070	19,750,759	1.46	111.97
1	TF32	0.019	0.020	0.020	0.020	8,912,296	-	404.19
4	TF32	0.054	0.055	0.055	0.055	12,471,624	-	141.40
8	TF32	0.100	0.100	0.100	0.101	13,543,317	-	76.78

FastPitch + WaveGlow (TorchScript, denoising)

Batch size	Precision	Avg latency (s)	Latency tolerance interval 90% (s)	Latency tolerance interval 95% (s)	Latency tolerance interval 99% (s)	Throughput (samples/sec)	Speed-up with mixed precision	Avg RTF
1	FP16	0.076	0.077	0.077	0.078	2,223,336	1.38	100.83
4	FP16	0.265	0.267	0.267	0.267	2,552,577	1.36	28.94
8	FP16	0.515	0.515	0.516	0.516	2,630,328	1.37	14.91
1	TF32	0.105	0.106	0.106	0.107	1,610,266	-	73.03
4	TF32	0.362	0.363	0.363	0.363	1,872,327	-	21.23
8	TF32	0.708	0.709	0.709	0.709	1,915,577	-	10.86

Inference performance: NVIDIA DGX-1 (1x V100 16GB)

Our results were obtained by running the ./scripts/inference_benchmark.sh script in the PyTorch 22.08-py3 NGC container. The input utterance has 128 characters, synthesized audio has 8.05 s.

FastPitch (TorchScript, denoising)

Batch size	Precision	Avg latency (s)	Latency tolerance interval 90% (s)	Latency tolerance interval 95% (s)	Latency tolerance interval 99% (s)	Throughput (frames/sec)	Speed-up with mixed precision	Avg RTF
1	FP16	0.007	0.008	0.008	0.008	88,908	1.10	1032.23
4	FP16	0.010	0.010	0.010	0.010	272,564	1.73	791.12
8	FP16	0.013	0.013	0.013	0.013	415,263	2.35	602.65
1	FP32	0.008	0.008	0.008	0.009	80,558	-	935.28
4	FP32	0.017	0.017	0.017	0.017	157,114	-	456.02
8	FP32	0.030	0.030	0.030	0.030	176,754	-	256.51

FastPitch + HiFi-GAN (TorchScript, denoising)

Batch size	Precision	Avg latency (s)	Latency tolerance interval 90% (s)	Latency tolerance interval 95% (s)	Latency tolerance interval 99% (s)	Throughput (samples/sec)	Speed-up with mixed precision	Avg RTF
1	FP16	0.025	0.025	0.025	0.025	6,788,274	2.09	307.86
4	FP16	0.067	0.068	0.068	0.068	10,066,291	2.63	114.13
8	FP16	0.123	0.124	0.124	0.124	10,992,774	2.78	62.32
1	FP32	0.052	0.053	0.053	0.053	3,246,699	-	147.24
4	FP32	0.177	0.178	0.179	0.179	3,829,018	-	43.41
8	FP32	0.343	0.345	0.345	0.346	3,953,920	-	22.41

FastPitch + WaveGlow (TorchScript, denoising)

Batch size	Precision	Avg latency (s)	Latency tolerance interval 90% (s)	Latency tolerance interval 95% (s)	Latency tolerance interval 99% (s)	Throughput (samples/sec)	Speed-up with mixed precision	Avg RTF
1	FP16	0.134	0.135	0.135	0.135	1,259,550	2.89	57.12
4	FP16	0.503	0.504	0.505	0.505	1,346,145	2.88	15.26
8	FP16	0.995	0.999	0.999	1.001	1,360,952	2.89	7.72
1	FP32	0.389	0.391	0.392	0.393	435,564	-	19.75
4	FP32	1.453	1.455	1.456	1.457	466,685	-	5.29
8	FP32	2.875	2.879	2.880	2.882	471,602	-	2.67

Inference performance: NVIDIA T4

Our results were obtained by running the ./scripts/inference_benchmark.sh script in the PyTorch 22.08-py3 NGC container. The input utterance has 128 characters, synthesized audio has 8.05 s.

FastPitch (TorchScript, denoising)

Batch size	Precision	Avg latency (s)	Latency tolerance interval 90% (s)	Latency tolerance interval 95% (s)	Latency tolerance interval 99% (s)	Throughput (frames/sec)	Speed-up with mixed precision	Avg RTF
1	FP16	0.008	0.008	0.008	0.008	87,937	1.69	1020.95
4	FP16	0.017	0.017	0.017	0.018	154,880	2.55	449.54
8	FP16	0.029	0.030	0.030	0.030	181,776	2.61	263.80
1	FP32	0.013	0.013	0.013	0.013	52,062	-	604.45
4	FP32	0.044	0.045	0.045	0.045	60,733	-	176.28
8	FP32	0.076	0.077	0.077	0.077	69,685	-	101.13

FastPitch + HiFi-GAN (TorchScript, denoising)

Batch size	Precision	Avg latency (s)	Latency tolerance interval 90% (s)	Latency tolerance interval 95% (s)	Latency tolerance interval 99% (s)	Throughput (samples/sec)	Speed-up with mixed precision	Avg RTF
1	FP16	0.055	0.056	0.056	0.057	3,076,809	2.55	139.54
4	FP16	0.201	0.203	0.204	0.204	3,360,014	2.67	38.10
8	FP16	0.393	0.395	0.396	0.397	3,444,245	2.65	19.53
1	FP32	0.140	0.142	0.142	0.142	1,208,678	-	54.82
4	FP32	0.538	0.542	0.543	0.545	1,260,627	-	14.29
8	FP32	1.045	1.049	1.050	1.051	1,297,726	-	7.36

FastPitch + WaveGlow (TorchScript, denoising)

Batch size	Precision	Avg latency (s)	Latency tolerance interval 90% (s)	Latency tolerance interval 95% (s)	Latency tolerance interval 99% (s)	Throughput (samples/sec)	Speed-up with mixed precision	Avg RTF
1	FP16	0.409	0.411	0.411	0.412	414,019	2.65	18.78
4	FP16	1.619	1.622	1.623	1.624	418,010	2.91	4.74
8	FP16	3.214	3.219	3.220	3.222	421,148	2.72	2.39
1	FP32	1.084	1.087	1.088	1.089	156,345	-	7.09
4	FP32	4.721	4.735	4.738	4.743	143,585	-	1.63
8	FP32	8.764	8.777	8.779	8.784	154,694	-	0.88

Release notes

We're constantly refining and improving our performance on AI and HPC workloads even on the same hardware, with frequent updates to our software stack. For our latest performance data, refer to these pages for AI and HPC benchmarks.

Changelog

October 2022

Updated performance tables

July 2022

Performance optimizations, speedups up to 1.2x (DGX-1) and 1.6x (DGX A100)

June 2022

MHA bug fix affecting models with > 1 attention heads

August 2021

Improved quality of synthesized audio
Added capability to automatically align audio to transcripts during training without a pre-trained Tacotron 2 aligning model
Added capability to train on both graphemes and phonemes
Added conditioning on energy
Faster training recipe
F0 is now estimated with Probabilistic YIN (PYIN)
Updated performance tables
Changed version of FastPitch from 1.0 to 1.1

October 2020

Added multispeaker capabilities
Updated text processing module

June 2020

Updated performance tables to include A100 results

May 2020

Initial release

Known issues

There are no known issues with this model.