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.

• Model overview
  • Model architecture
  • Default configuration
  • Feature support matrix
    • Features
  • Mixed precision training
    • Enabling mixed precision
    • Enabling TF32
  • Glossary
• Setup
  • Requirements
• Quick Start Guide
• Advanced
  • Scripts and sample code
  • Parameters
  • Command-line options
  • Getting the data
    • Dataset guidelines
    • Multi-dataset
  • Training process
  • Inference process
  • Example: Training a model on Mandarin Chinese
• Performance
  • Benchmarking
    • Training performance benchmark
    • Inference performance benchmark
  • Results
    • Training accuracy results
      • Training accuracy: NVIDIA DGX A100 (8x A100 80GB)
      • Training accuracy: NVIDIA DGX-1 (8x V100 16GB)
    • Training performance results
      • Training performance: NVIDIA DGX A100 (8x A100 80GB)
      • Training performance: NVIDIA DGX-1 (8x V100 16GB)
      • Expected training time
    • Inference performance results
      • Inference performance: NVIDIA DGX A100 (1x A100 80GB)
      • Inference performance: NVIDIA DGX-1 (1x V100 16GB)
      • Inference performance: NVIDIA T4
• Release notes
  • Changelog
  • Known issues

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).

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

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.

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.

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.

The following features are supported by this model.

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 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:

1. Porting the model to use the FP16 data type where appropriate.
2. 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.

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

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.

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.

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

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:
  • NVIDIA Volta architecture
  • NVIDIA Turing architecture
  • NVIDIA Ampere architecture

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:

• Getting Started Using NVIDIA GPU Cloud
• Accessing And Pulling From The NGC Container Registry
• Running PyTorch

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.

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.

1. Clone the repository.

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

2. 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

3. 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

4. Start training.

   bash scripts/train.sh

   The training will produce a FastPitch model capable of generating mel-spectrograms from raw text. It will be serialized as a single .pt checkpoint file, along with a series of intermediate checkpoints. The script is configured for 8x GPU with at least 16GB of memory. Consult Training process and example configs to adjust to a different configuration or enable Automatic Mixed Precision.

5. Start validation/evaluation.

   Ensure your training loss values are comparable to those listed in the table in the Results section. Note that the validation loss is evaluated with ground truth durations for letters (not the predicted ones). The loss values are stored in the ./output/nvlog.json log file, ./output/{train,val,test} as TensorBoard logs, and printed to the standard output (stdout) during training. The main reported loss is a weighted sum of losses for mel-, pitch-, and duration- predicting modules.

   The audio can be generated by following the Inference process section below. The synthesized audio should be similar to the samples in the ./audio directory.

6. Start inference/predictions.

   To synthesize audio, you will need a WaveGlow model, which generates waveforms based on mel-spectrograms generated with FastPitch. By now, a pre-trained model should have been downloaded by the scripts/download_dataset.sh script. Alternatively, to train WaveGlow from scratch, follow the instructions in NVIDIA/DeepLearningExamples/Tacotron2 and replace the checkpoint in the ./pretrained_models/waveglow directory.

   You can perform inference using the respective .pt checkpoints that are passed as --fastpitch and --waveglow arguments:

   python inference.py \
       --cuda \
       --fastpitch output/<FastPitch checkpoint> \
       --energy-conditioning \
       --waveglow pretrained_models/waveglow/<WaveGlow checkpoint> \
       --wn-channels 256 \
       -i phrases/devset10.tsv \
       -o output/wavs_devset10

   The speech is generated from a file passed with the -i argument, with one utterance per line:

   `<output wav file name>|<utterance>`

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.

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

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.

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)

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

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.

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.

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

1. Prepare a directory with .wav files.

   ./my_dataset
   └── wavs

2. 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.

3. 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

4. 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

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

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.

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.

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).

1. Set up the repository and run a Docker container

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

2. 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_*.

3. 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.

4. 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.

5. 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.

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

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.

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.

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

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.

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.

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.

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

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.

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

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

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

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.

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)

FastPitch + HiFi-GAN (TorchScript, denoising)

FastPitch + WaveGlow (TorchScript, denoising)

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)

FastPitch + HiFi-GAN (TorchScript, denoising)

FastPitch + WaveGlow (TorchScript, denoising)

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)

FastPitch + HiFi-GAN (TorchScript, denoising)

FastPitch + WaveGlow (TorchScript, denoising)

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.

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

There are no known issues with this model.