DataBloomGrowing up in a military family, Christopher Scott moved more than 30 times, which instilled in him “the ability to be comfortable with, and even motivated by, new environments,” he said.
The post Meet the Omnivore: Christopher Scott Constructs Architectural Designs, Virtual Environments With NVIDIA Omniverse appeared first on NVIDIA Blog.
Learn about new CUDA features, digital twins for weather and climate, quantum circuit simulations, and much more with these GTC 2022 sessions.
Learn about new CUDA features, digital twins for weather and climate, quantum circuit simulations, and much more with these GTC 2022 sessions.
There is an abundance of market-approved medical AI software that can be used to improve patient care and hospital operations, but we have not yet seen these…
There is an abundance of market-approved medical AI software that can be used to improve patient care and hospital operations, but we have not yet seen these technologies create the large-scale transformation in healthcare that was expected.
Adopting cutting-edge technologies is not a trivial exercise for healthcare institutions. It requires a balance of legal, clinical, and technical risks against the promise of improved patient outcomes and operational efficiency.
Traditionally, the challenges around the adoption of such technologies fell into one of three buckets: people, platforms, and policy. The challenges around a platform for AI adoption are particularly unique given the nature of deep learning technology and the current state of the medical AI ecosystem.
Most deep learning applications have a narrow field of scope. If they stray beyond their domain, they can exhibit unpredictable and unintuitive behavior. This means that to achieve large-scale transformation in medicine, we need thousands of AI applications.
Each of these AI models in production will be communicating information with live clinical systems and making all kinds of inferences that must then be managed. This has the potential of creating an “AI jungle,” with a huge amount of technical debt in an environment where there hasn’t been substantial investment in people to manage such risks.
Another challenge for deploying AI at scale is the lack of interoperability of AI models. Deployment and data integration don’t scale within or across institutions. This lack of interoperability exists within information systems and semantics and between organizations. The result is a high barrier of entry for data scientists and start-ups who don’t have the capacity or domain knowledge to make an impact.
Finally, in recognition of the immaturity of the medical AI economy relative to other subdomains in MedTech, evidence generation must be at the heart of the design of a platform for AI adoption, as many of the AI applications on the market today still require extensive research and analysis of their performance. This is true not only from the perspective of monitoring but also to measure impact on health outcomes.
An AI platform that solves these challenges must solve them at the enterprise level to fully capture the benefits of the ‘virtuous cycle of AI’ and to fully mitigate risks around deployment of artificial intelligence.
This ensures the lowered costs of deployment by plugging into a platform that is already integrated to the clinical information systems within healthcare facilities. It also lowers costs around staff and support by creating an opportunity for a single team to service the entire institution, empowered with an enterprise-wide view for managing risks and continuous improvement.
AIDE, developed by the UK Government-funded AI Centre for Value Based Healthcare, is a new operating system for the hospital that allows healthcare providers to deploy AI models safely, effectively, and efficiently. It provides a harmonized hardware and software layer that facilitates the deployment and use of any AI application.
There are numerous technical risks involved in deploying a large number of models. AIDE mitigates these by providing an administration view that reports every inference of every deployed model as well as performance trend analysis to enable real-time intervention in the case of poor performance.
AIDE also solves the challenge of interoperability by packaging and deploying containerized applications and communicating with the rest of the hospital through standard protocols such as DICOM, HL7, and FHIR.
Clinicians can also review AI inference results with AIDE before they are sent to the patient’s electronic health record (EHR). This clinical review stage can collect useful data around failure instances, which can be fed back to the developer and close the feedback loop.
When considering the wide scale adoption of AI, it is important to first consider, as an analogous example, the discovery of X-rays and the subsequent transformation of healthcare through the development of radiology.
After the discovery by Dr Wilhelm Roentgen in 1895 and the famous X-ray of his wife Bertha’s hand, the first uses of X-ray technology were for industrial applications, such as welding inspection, and consumer applications, such as shoe-fitting, rather than medical applications.
Today, most patient experiences involve medical imaging for diagnosis, prognosis, treatment monitoring and more. Rural medical centers can acquire images in the middle of the night and have them reported within an hour by a specialist in another part of the world.
That kind of transformation was only made possible almost 100 years after the invention of the x-ray when the American College of Radiology and National Electrical Manufacturers Association published a standard for the encoding and transfer of medical images, named “Digital Imaging and Communications in Medicine.”
With the birth of the standard in the early 1990s, a transformational journey had begun that would change what would be possible in the art of medicine, leading to advances in oncology, neurology, and many other medical specialties.
Similarly, with deep learning, industrial and consumer applications have raced ahead while medical applications have had limited adoption and even less transformational impact.
That is why the key innovation in AIDE, as an enterprise AI platform, is that it is built on top of the open-source MONAI Deploy architecture.
MONAI Deploy was built to bridge the gap from research innovation to validation and clinical production environments. It gives developers and researchers a default standard, called MONAI Deploy Application Package (MAP), that easily integrates into health IT standards such as DICOM. It also integrates into deployment options across a variety of data center, cloud, and edge environments, making it easy for you to adopt new medical AI applications.
The MONAI Deploy Working Group has defined an open architecture and standard APIs for developing, packaging, testing, deploying, and running medical AI applications in clinical production.
The high-level architecture includes the following components:
The system has been designed to allow pluggable execution of tasks by different inference engines. The MONAI community is going to keep moving in that direction as well.
The MONAI Deploy architecture has been co-designed by an international community of hardware, software, academic, and healthcare partners for the mutual aim of standardizing the medical AI lifecycle. This is much in the same way the ACR and NEMA did with medical images three decades ago.
This new layer of informatics, built on top of existing clinical information systems and medical devices, will help usher in the new era of data-driven medicine. For more information about AIDE, see AI Centre for Value Based Healthcare Platforms.
TGIGFNT: thank goodness it’s GFN Thursday. Start your weekend early with seven new games joining the GeForce NOW library of over 1,400 titles. Whether it’s streaming on an older-than-the-dinosaurs PC, a Mac that normally couldn’t dream of playing PC titles, or mobile devices – it’s all possible to play your way thanks to GeForce NOW. Read article >
The post GFN Thursday Delivers Seven New Games This Week appeared first on NVIDIA Blog.
Join us for these featured GTC 2022 sessions to learn about optimizing PyTorch models, accelerating graph neural networks, improving GPU performance with…
Join us for these featured GTC 2022 sessions to learn about optimizing PyTorch models, accelerating graph neural networks, improving GPU performance with automated code generation, and more.
An exaflop is a measure of performance for a supercomputer that can calculate at least one quintillion floating point operations per second.
An exaflop is a measure of performance for a supercomputer that can calculate at least one quintillion floating point operations per second.
As consumers expect faster, cheaper deliveries, companies are turning to AI to rethink how they move goods. Foremost among these new systems are “hub-and-spoke,” or middle-mile, operations, where companies place distribution centers closer to retail operations for quicker access to inventory. However, faster delivery is just part of the equation. These systems must also be Read article >
The post Reinventing the Wheel: Gatik’s Apeksha Kumavat Accelerates Autonomous Delivery for Wal-Mart and More appeared first on NVIDIA Blog.
Virtual reality (VR), augmented reality (AR), and mixed reality (MR) environments can feel incredibly real due to the physically immersive experience. Adding a…
Virtual reality (VR), augmented reality (AR), and mixed reality (MR) environments can feel incredibly real due to the physically immersive experience. Adding a voice-based interface to your extended reality (XR) application can make it appear even more realistic.
Imagine using your voice to navigate through an environment or giving a verbal command and hearing a response back from a virtual entity.
The possibilities to harness speech AI in XR environments is fascinating. Speech AI skills, such as automatic speech recognition (ASR) and text-to-speech (TTS), make XR applications enjoyable, easy to use, and more accessible to users with speech impairments.
This post explains how speech recognition, also referred to as speech-to-text (STT), can be used in your XR app, what ASR customizations are available, and how to get started with running ASR services in your Windows applications.
In most of today’s XR experiences, users don’t have access to a keyboard or mouse. The way VR game controllers typically interact with a virtual experience is clumsy and unintuitive, making navigation through menus difficult when you’re immersed in the environment.
When virtually immersed, we want our experience to feel natural, both in how we perceive it and in how we interact with it. Speech is one of the most common interactions that we use in the real world.
Adding speech AI-enabled voice commands and responses to your XR application makes interaction feel much more natural and dramatically simplifies the learning curve for users.
Today, there are a wide array of wearable tech devices that enable people to experience immersive realities while using their voice:
VR can help businesses save costs by automating a number of tasks in the automotive industry, such as modeling cars, training assembly workers, and driving simulations.
An added speech AI component makes hands-free interactions possible. For example, users can leverage STT skills to give commands to VR apps, and apps can respond in a way that sounds human with TTS.
As shown in Figure 1, a user sends an audio request to a VR application that is then converted to text using ASR. Natural language understanding takes text as an input and generates a response, which is spoken back to the user using TTS.
Developing speech AI pipelines is not as easy as it sounds. Traditionally, there has always been a trade-off between accuracy and real-time response when building pipelines.
This post focuses solely on ASR, and we examine some of today’s available customizations for XR app developers. We also discuss using NVIDIA Riva, a GPU-accelerated speech AI SDK, for building applications customized for specific use cases while delivering real-time performance.
An ASR pipeline includes a feature extractor, acoustic model, decoder or language model, and punctuation and capitalization model (Figure 2).
To understand the ASR customizations available, it’s important to grasp the end-to-end process. First, feature extraction takes place to turn raw audio waveforms into spectrograms / mel spectrograms. These spectrograms are then fed into an acoustic model that generates a matrix with probabilities for all the characters at each time step.
Next, the decoder, in conjunction with the language model, uses that matrix as an input to produce a transcript. You can then run the resulting transcript through the punctuation and capitalization model to improve readability.
Advanced speech AI SDKs and workflows, such as Riva, support speech recognition pipeline customization. Customization helps you address several language-specific challenges, such as understanding one or more of the following:
Customizations in Riva can be applied in both the training and inference stages. Starting with training-level customizations, you can fine-tune acoustic models, decoder/language models, and punctuation and capitalization models. This ensures that your pipeline understands different language, dialects, accents, and industry-specific jargon, and is robust to noise.
When it comes to inference-level customizations, you can use word boosting. With word boosting, the ASR pipeline is more likely to recognize certain words of interest by giving them a higher score when decoding the output of the acoustic model.
Riva runs as a client-server model. To run Riva, you need access to a Linux server with an NVIDIA GPU, where you can install and run the Riva server (specifics and instructions are provided in this post).
The Riva client API is integrated into your Windows application. At runtime, the Windows client sends Riva requests over the network to the Riva server, and the Riva server sends back replies. A single Riva server can simultaneously support many Riva clients.
ASR services can be run in two different modes:
Examples showing both modes are provided later in this post.
In this section, you learn several ways to integrate Riva into your Windows application:
First, here’s how to set up and run the Riva server.
ngc commands from a command-line interface (CLI).nvidia-docker package.Download the scripts from NGC by running the following command:
ngc registry resource download-version nvidia/riva/riva_quickstart:2.4.0
Initialize the Riva server:
bash riva_init.sh
Start the Riva server:
bash riva_start.sh
For more information about the latest updates, see the NVIDIA Riva Quick Start Guide.
First, run the following command to install the riva client package. Make sure that you’re using Python version 3.7.
pip install nvidia-riva-client
The following code example runs ASR transcription in offline mode. You must change the server address, give the path to the audio file to be transcribed, and select the language code for your choice. Currently, Riva supports English, Spanish, German, Russian, and Mandarin.
import io
import IPython.display as ipd
import grpc
import riva.client
auth = riva.client.Auth(uri='server address:port number')
riva_asr = riva.client.ASRService(auth)
# Supports .wav file in LINEAR_PCM encoding, including .alaw, .mulaw, and .flac formats with single channel
# read in an audio file from local disk
path = "audio file path"
with io.open(path, 'rb') as fh:
content = fh.read()
ipd.Audio(path)
# Set up an offline/batch recognition request
config = riva.client.RecognitionConfig()
#req.config.encoding = ra.AudioEncoding.LINEAR_PCM # Audio encoding can be detected from wav
#req.config.sample_rate_hertz = 0 # Sample rate can be detected from wav and resampled if needed
config.language_code = "en-US" # Language code of the audio clip
config.max_alternatives = 1 # How many top-N hypotheses to return
config.enable_automatic_punctuation = True # Add punctuation when end of VAD detected
config.audio_channel_count = 1 # Mono channel
response = riva_asr.offline_recognize(content, config)
asr_best_transcript = response.results[0].alternatives[0].transcript
print("ASR Transcript:", asr_best_transcript)
print("nnFull Response Message:")
print(response)
To run an ASR streaming client, clone the riva python-clients repository and run the file that comes with the repository.
To get the ASR streaming client to work on Windows, clone the repository by running the following command:
git clone https://github.com/nvidia-riva/python-clients.git
From the python-clients/scripts/asr folder, run the following command:
python transcribe_mic.py --server=server address:port number
Here’s the transcribe_mic.py:
import argparse
import riva.client
from riva.client.argparse_utils import add_asr_config_argparse_parameters, add_connection_argparse_parameters
import riva.client.audio_io
def parse_args() -> argparse.Namespace:
default_device_info = riva.client.audio_io.get_default_input_device_info()
default_device_index = None if default_device_info is None else default_device_info['index']
parser = argparse.ArgumentParser(
description="Streaming transcription from microphone via Riva AI Services",
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument("--input-device", type=int, default=default_device_index, help="An input audio device to use.")
parser.add_argument("--list-devices", action="store_true", help="List input audio device indices.")
parser = add_asr_config_argparse_parameters(parser, profanity_filter=True)
parser = add_connection_argparse_parameters(parser)
parser.add_argument(
"--sample-rate-hz",
type=int,
help="A number of frames per second in audio streamed from a microphone.",
default=16000,
)
parser.add_argument(
"--file-streaming-chunk",
type=int,
default=1600,
help="A maximum number of frames in a audio chunk sent to server.",
)
args = parser.parse_args()
return args
def main() -> None:
args = parse_args()
if args.list_devices:
riva.client.audio_io.list_input_devices()
return
auth = riva.client.Auth(args.ssl_cert, args.use_ssl, args.server)
asr_service = riva.client.ASRService(auth)
config = riva.client.StreamingRecognitionConfig(
config=riva.client.RecognitionConfig(
encoding=riva.client.AudioEncoding.LINEAR_PCM,
language_code=args.language_code,
max_alternatives=1,
profanity_filter=args.profanity_filter,
enable_automatic_punctuation=args.automatic_punctuation,
verbatim_transcripts=not args.no_verbatim_transcripts,
sample_rate_hertz=args.sample_rate_hz,
audio_channel_count=1,
),
interim_results=True,
)
riva.client.add_word_boosting_to_config(config, args.boosted_lm_words, args.boosted_lm_score)
with riva.client.audio_io.MicrophoneStream(
args.sample_rate_hz,
args.file_streaming_chunk,
device=args.input_device,
) as audio_chunk_iterator:
riva.client.print_streaming(
responses=asr_service.streaming_response_generator(
audio_chunks=audio_chunk_iterator,
streaming_config=config,
),
show_intermediate=True,
)
if __name__ == '__main__':
main()
Here’s how to run a Riva ASR offline client using Docker in C++.
Clone the /cpp-clients GitHub repository by running the following command:
git clone https://github.com/nvidia-riva/cpp-clients.git
Build the Docker image:
DOCKER_BUILDKIT=1 docker build . –tag riva-client
Run the Docker image:
docker run -it --net=host riva-client
Start the Riva speech recognition client:
Riva_asr_client –riva_url server address:port number –audio_file audio_sample
To run the ASR streaming client riva_asr in C++, you must first compile the cpp sample. It is straightforward using CMake, after the following dependencies are met:
gflagsgloggrpcrtaudiorapidjsonprotobufgrpc_cpp_pluginCreate a folder /build within the root source folder. From the terminal, type cmake .. and then make. For more information, see the readme file included in the repository.
After the sample is compiled, run it by entering the following command:
riva_asr.exe --riva_uri={riva server url}:{riva server port} --audio_device={Input device name, e.g. "plughw:PCH,0"}
riva_uri: The address:port value of the riva server. By default, the riva server listens to port 50051.audio_device: The input device (microphone) to be used.The sample implements essentially four steps. Only a few short examples are shown in this post. For more information, see the file streaming_recognize_client.cc.
Open the input stream using the input (microphone) device specified from the command line. In this case, you are using one channel at 16K samples per second and 16 bits per sample.
int StreamingRecognizeClient::DoStreamingFromMicrophone(const std::string& audio_device, bool& request_exit)
{
nr::AudioEncoding encoding = nr::LINEAR_PCM;
adc.setErrorCallback(rtErrorCallback);
RtAudio::StreamParameters parameters;
parameters.nChannels = 1;
parameters.firstChannel = 0;
unsigned int sampleRate = 16000;
unsigned int bufferFrames = 1600; // (0.1 sec of rec) sample frames
RtAudio::StreamOptions streamOptions;
streamOptions.flags = RTAUDIO_MINIMIZE_LATENCY;
…
RtAudioErrorType error = adc.openStream( nullptr, ¶meters, RTAUDIO_SINT16, sampleRate, &bufferFrames, &MicrophoneCallbackMain, static_cast(&uData), &streamOptions);
Open the grpc communication channel with the Riva server using the protocol api interface specified by the .proto files (in the source in the folder riva/proto):
int StreamingRecognizeClient::DoStreamingFromMicrophone(const std::string& audio_device, bool& request_exit)
{
…
std::shared_ptr call = std::make_shared(1, word_time_offsets_);
call->streamer = stub_->StreamingRecognize(&call->context);
// Send first request
nr_asr::StreamingRecognizeRequest request;
auto streaming_config = request.mutable_streaming_config();
streaming_config->set_interim_results(interim_results_);
auto config = streaming_config->mutable_config();
config->set_sample_rate_hertz(sampleRate);
config->set_language_code(language_code_);
config->set_encoding(encoding);
config->set_max_alternatives(max_alternatives_);
config->set_audio_channel_count(parameters.nChannels);
config->set_enable_word_time_offsets(word_time_offsets_);
config->set_enable_automatic_punctuation(automatic_punctuation_);
config->set_enable_separate_recognition_per_channel(separate_recognition_per_channel_);
config->set_verbatim_transcripts(verbatim_transcripts_);
if (model_name_ != "") {
config->set_model(model_name_);
}
call->streamer->Write(request);
Start sending audio data, received by the microphone to riva through grpc messages:
static int MicrophoneCallbackMain( void *outputBuffer, void *inputBuffer, unsigned int nBufferFrames, double streamTime, RtAudioStreamStatus status, void *userData )
Receive the transcribed audio through grpc answers from the server:
void
StreamingRecognizeClient::ReceiveResponses(std::shared_ptr call, bool audio_device)
{
…
while (call->streamer->Read(&call->response)) { // Returns false when no m ore to read.
call->recv_times.push_back(std::chrono::steady_clock::now());
// Reset the partial transcript
call->latest_result_.partial_transcript = "";
call->latest_result_.partial_time_stamps.clear();
bool is_final = false;
for (int r = 0; r response.results_size(); ++r) {
const auto& result = call->response.results(r);
if (result.is_final()) {
is_final = true;
}
…
call->latest_result_.audio_processed = result.audio_processed();
if (print_transcripts_) {
call->AppendResult(result);
}
}
if (call->response.results_size() && interim_results_ && print_transcripts_) {
std::cout latest_result_.final_transcripts[0] +
call->latest_result_.partial_transcript
recv_final_flags.push_back(is_final);
}
By recognizing your voice or carrying out a command, speech AI is expanding from empowering actual humans in contact centers to empowering digital humans in the metaverse.
For more information about how to add speech AI skills to your applications, see the following resources
For more information about how businesses are deploying Riva in production to improve their services, see the customer stories in Solution Showcase.
AI processing requires full-stack innovation across hardware and software platforms to address the growing computational demands of neural networks. A key area…
AI processing requires full-stack innovation across hardware and software platforms to address the growing computational demands of neural networks. A key area to drive efficiency is using lower precision number formats to improve computational efficiency, reduce memory usage, and optimize for interconnect bandwidth.
To realize these benefits, the industry has moved from 32-bit precisions to 16-bit, and now even 8-bit precision formats. Transformer networks, which are one of the most important innovations in AI, benefit from an 8-bit floating point precision in particular. We believe that having a common interchange format will enable rapid advancements and the interoperability of both hardware and software platforms to advance computing.
NVIDIA, Arm, and Intel have jointly authored a whitepaper, FP8 Formats for Deep Learning, describing an 8-bit floating point (FP8) specification. It provides a common format that accelerates AI development by optimizing memory usage and works for both AI training and inference. This FP8 specification has two variants, E5M2 and E4M3.
This format is natively implemented in the NVIDIA Hopper architecture and has shown excellent results in initial testing. It will immediately benefit from the work being done by the broader ecosystem, including the AI frameworks, in implementing it for developers.
FP8 minimizes deviations from existing IEEE 754 floating point formats with a good balance between hardware and software to leverage existing implementations, accelerate adoption, and improve developer productivity.
E5M2 uses five bits for the exponent and two bits for the mantissa and is a truncated IEEE FP16 format. In circumstances where more precision is required at the expense of some numerical range, the E4M3 format makes a few adjustments to extend the range representable with a four-bit exponent and a three-bit mantissa.
The new format saves additional computational cycles since it uses just eight bits. It can be used for both AI training and inference without requiring any re-casting between precisions. Furthermore, by minimizing deviations from existing floating point formats, it enables the greatest latitude for future AI innovation while still adhering to current conventions.
Testing the proposed FP8 format shows comparable accuracy to 16-bit precisions across a wide array of use cases, architectures, and networks. Results on transformers, computer vision, and GAN networks all show that FP8 training accuracy is similar to 16-bit precisions while delivering significant speedups. For more information about accuracy studies, see the FP8 Formats for Deep Learning whitepaper.
In Figure 1, different networks use different accuracy metrics (PPL and Loss), as indicated.
In MLPerf Inference v2.1, the AI industry’s leading benchmark, NVIDIA Hopper leveraged this new FP8 format to deliver a 4.5x speedup on the BERT high-accuracy model, gaining throughput without compromising on accuracy.
NVIDIA, Arm, and Intel have published this specification in an open, license-free format to encourage broad industry adoption. They will also submit this proposal to IEEE.
By adopting an interchangeable format that maintains accuracy, AI models will operate consistently and performantly across all hardware platforms, and help advance the state of the art of AI.
Standards bodies and the industry as a whole are encouraged to build platforms that can efficiently adopt the new standard. This will help accelerate AI development and deployment by providing a universal, interchangeable precision.
Think fast. Enterprise AI, new gaming technology, the metaverse and the 3D internet, and advanced AI technologies tailored to just about every industry are all coming your way. NVIDIA founder and CEO Jensen Huang’s keynote at NVIDIA GTC on Tuesday, Sept. 20, is the best way to get ahead of all these trends. NVIDIA’s virtual Read article >
The post Get up to Speed: Five Reasons Not to Miss NVIDIA CEO Jensen Huang’s GTC Keynote Sept. 20 appeared first on NVIDIA Blog.