Category: Misc

NVIDIA today introduced NVIDIA DRIVE™ Thor, its next-generation centralized computer for safe and secure autonomous vehicles.

NVIDIA today announced its first software- and infrastructure-as-a-service offering — NVIDIA Omniverse™ Cloud — a comprehensive suite of cloud services for artists, developers and enterprise teams to design, publish, operate and experience metaverse applications anywhere.

Deutsche Bahn’s rail network consists of 5,700 stations and 33,000 kilometers of track, making it the largest in Western Europe. Digitale Schiene Deutschland (Digital Rail for Germany, or DSD), part of Germany’s national railway operator Deutsche Bahn, is working to increase the network’s capacity without building new tracks. It’s striving to create a powerful railway Read article >

The post On Track: Digitale Schiene Deutschland Building Digital Twin of Rail Network in NVIDIA Omniverse appeared first on NVIDIA Blog.

With tens of millions of weekly transactions across its more than 2,000 stores, Lowe’s helps customers achieve their home-improvement goals. Now, the Fortune 50 retailer is experimenting with high-tech methods to elevate both the associate and customer experience. Using NVIDIA Omniverse Enterprise to visualize and interact with a store’s digital data, Lowe’s is testing digital Read article >

The post Reinventing Retail: Lowe’s Teams With NVIDIA and Magic Leap to Create Interactive Store Digital Twins appeared first on NVIDIA Blog.

The latest NVIDIA HPC SDK update expands portability and now supports the Arm-based AWS Graviton3 processor. In this post, you learn how to enable Scalable…

The latest NVIDIA HPC SDK update expands portability and now supports the Arm-based AWS Graviton3 processor. In this post, you learn how to enable Scalable Vector Extension (SVE) auto-vectorization with the NVIDIA compilers to maximize the performance of HPC applications running on the AWS Graviton3 CPU.

NVIDIA HPC SDK

The NVIDIA HPC SDK includes the proven compilers, libraries, and software tools essential to maximizing developer productivity and building HPC applications for GPUs, CPUs, or the cloud.

NVIDIA HPC compilers enable cross-platform C, C++, and Fortran programming for NVIDIA GPUs and multicore Arm, OpenPOWER, or x86-64 CPUs. These are ideal for HPC modeling and simulation applications written in C, C++, or Fortran with OpenMP, OpenACC, and CUDA. 

For example, SPEC CPU® 2017 benchmark scores are estimated to increase by 17% on the AWS Graviton 3 when compiled with the NVIDIA HPC compilers vs. GCC 12.1.

The compilers are also fully interoperable with the optimized NVIDIA math libraries, communication libraries, and performance tuning and debugging tools. These accelerated math libraries maximize performance on common HPC algorithms, and the optimized communications libraries enable standards-based scalable systems programming.

The integrated performance profiling and debugging tools simplify porting and optimization of HPC applications, and the containerization tools enable easy deployment on-premises or in the cloud.

Arm and AWS Graviton3

AWS Graviton3 launched in May 2022 as the Arm-based CPU from AWS. The Arm architecture has a legacy of power efficiency and support for high memory bandwidth that makes it ideal for cloud and data center computing. Amazon reports:

The Amazon EC2 C7g instances, powered by the latest generation AWS Graviton3 processors, provide the best price performance in Amazon EC2 for compute-intensive workloads. C7g instances are ideal for HPC, batch processing, electronic design automation (EDA), gaming, video encoding, scientific modeling, distributed analytics, CPU-based machine learning (ML) inference, and ad-serving. They offer up to 25% better performance over the sixth generation AWS Graviton2-based C6g instances.

Compared to AWS Graviton2, ANSYS benchmarked 35% better performance on AWS Graviton3. Formula 1 simulations are also 40% faster. Arm-based CPUs have been delivering significant innovations and performance enhancements since the launch of the Arm Neoverse product line, when the Neoverse N1 core exceeded performance expectations by 30%.

In keeping with the history of Arm enabling support for new computing technologies well ahead of the competition, AWS Graviton3 features DDR5 memory and the SVE to the Arm architecture.

Amazon EC2 C7g instances are the first in the cloud to feature DDR5 memory, which provides 50% higher memory bandwidth compared to DDR4 memory to enable high-speed access to data in memory. The best way to take full advantage of all that memory bandwidth is to use the latest in vectorization technologies: Arm SVE.

SVE architecture

In addition to being the first cloud-hosted CPU to offer DDR5, AWS Graviton3 is also the first in the cloud to feature SVE.

SVE was first introduced in the Fujitsu A64FX CPU, which powers the RIKEN Fugaku supercomputer. When Fugaku launched, it shattered all contemporary HPC CPU benchmarks and placed confidently at the top of the TOP500 supercomputers list for two years.

SVE and high-bandwidth memory are the key design features of the A64FX that make it ideal for HPC, and both these features are present in the AWS Graviton3 processor.

SVE is a next-generation SIMD extension to the Arm architecture. It enables flexible vector length implementations with a range of possible values in CPU implementations. The vector length can vary from a minimum of 128 bits to a maximum of 2,048 bits, at 128-bit increments.

For example, the Fujitsu A64FX implements SVE at 512-bits, while AWS Graviton3 implements it at 256-bits. Unlike other SIMD architectures, the same assembly code runs on both CPUs, even though the hardware vector bit-width is different. This is called vector-length agnostic (VLA) programming.

VLA code is highly portable and can enable compilers to generate better assembly code. But, if a compiler knows the target CPU’s hardware vector bit-width, it can enable further optimizations for that specific architecture. This is vector length–specific (VLS) programming.

SVE uses the same assembly language for both VLA and VLS. The only difference is that the compiler is free to make additional assertions about data layout, loop trip counts, and other relevant features while generating the code. This results in highly optimized, target-specific code that takes full advantage of the CPU.

SVE also introduces a powerful range of advanced features ideal for HPC and ML applications:

• Gather-load and scatter-store instructions allow operations on arrays-of-structures and other noncontiguous data to vectorize.
• Speculative vectorization enables the SIMD acceleration of string manipulation functions and loops that contain control flow.
• Horizontal and serialized vector operations facilitate data reductions and help optimize loops processing large datasets.

SVE is not an extension or the replacement of the NEON instruction set, which is also available in AWS Gravition3. SVE is redesigned for better data parallelism for HPC and ML.

Maximizing Graviton3 performance with NVIDIA HPC compilers

Compiler auto-vectorization is one of the easiest ways to take advantage of SVE, and the NVIDIA HPC compilers add support for SVE auto-vectorization in the 22.7 release.

To maximize performance, the compiler performs analysis to determine which SIMD instructions to generate. SVE auto-vectorization uses target-specific information to generate highly optimized vector length–specific (VLS) code based on the vector bit-width of the CPU core.

To enable SVE auto-vectorization, specify the appropriate -tp architecture flag for the target CPU: -tp=neoverse-v1. Not specifying a -tp option assumes that the application will be executed on the same system on which it was compiled.

Applications compiled with the NVIDIA HPC compilers on Graviton3 automatically take full advantage of the CPU’s 256-bit SVE SIMD units. Graviton3 is also backward compatible with the -tp=neoverse-n1 option but only runs vector code on its 128-bit NEON SIMD units.

Getting started with the NVIDIA HPC SDK

The NVIDIA HPC SDK provides a comprehensive and proven software stack. It enables HPC developers to create and optimize application performance on high-performance systems such as the NVIDIA platform and AWS Graviton3.

By providing a wide range of programming models, libraries, and development tools, applications can be efficiently developed for the specialized hardware that enables state-of-the-art performance in systems such as NVIDIA GPUs and SVE-enabled processors like AWS Graviton3.

For more information, see the following resources:

• Learn more about the compiler support and other posts at HPC SDK.
• Download the HPC SDK software for free.

An AI model card is a document that details how machine learning (ML) models work. Model cards provide detailed information about the ML model’s metadata…

An AI model card is a document that details how machine learning (ML) models work. Model cards provide detailed information about the ML model’s metadata including the datasets that it is based on, performance measures that it was trained on, and the deep learning training methodology itself. This post walks you through the current practice for AI model cards and how NVIDIA is planning to advance them with Model Card++, the enhanced next-generation AI model card. 

In their 2019 paper, Model Cards for Model Reporting, a group of data scientists, including Margaret Mitchell, Timnit Gebru, and Lucy Vasserman, sought to create a documentation standard for AI models. Their primary motivation was to promote transparency and accountability in the AI model development process by disclosing essential information about an AI model.

This information includes who developed the model, intended use cases and out-of-scope applications, expected users, how the model performs with different demographic groups, information about the data used to train and verify the model, limitations, and ethical considerations.

Until the development of the first AI model card, little information was shared about a particular AI model to help determine whether the model was suitable for a particular organization’s purpose. 

This becomes problematic if the output of a model could have an adverse impact on a particular group of people. For example, the 2019 university-led study, Discrimination through Optimization: How Facebook’s Ad Delivery Can Lead to Skewed Outcomes revealed that algorithms for delivering ads on social media resulted in discriminatory ad delivery despite the use of neutral parameters for targeting the ads.

The adoption of model cards helps the development and improvement of models by allowing developers to compare their results to those of similar models. Model cards highlight performance problems for those who plan to deploy a model. 

At the same time, model cards also educate policymakers who are drafting regulations and legislation governing AI models and systems. Although not required, developing model cards is a best practice that encourages developers to engage with the people who will ultimately be impacted by the model’s output.

The importance of model cards

While model cards are designed to encourage model transparency and trustworthiness, they are also used by stakeholders to improve developer understanding and standardize the decision-making processes.

Model cards are structured and organized in a schema. They concisely report information on different factors like demographics, environmental conditions, quantitative evaluation metrics, and, where provided, ethical considerations. 

Model cards can also record model version, type, date, license restrictions, information about the publishing organization, and other qualitative information. Model cards are designed to educate and allow an informed comparison of measures and benchmarks. Figure 1 shows the NGC Model Card for StyleGAN3.

Model cards are like open-source fact sheets. Unless you are the developer of the model itself, you would likely not even know much about the AI model itself without model cards. Model cards provide the most comprehensive understanding of a model’s details and considerations individuals should take into account for its application. 

For instance, a smartphone might have a face detection system that allows the user to unlock it based on recognition. Without model cards, model developers might not realize how a model will behave until it is deployed. This is what happened when Dr. Joy Buolamwini tried to use a face detection system as part of her graduate work at MIT.

AI model card accessibility

AI model cards should not just be built for developers; companies should also build model cards that are accessible to and readable by nontechnical individuals and technical experts alike. 

Model cards are not restricted to a given industry or domain. They can be used for computer vision, speech, recommender systems, and other AI workflows. In addition to having active use in higher education and research and high performance computing spaces, model cards have utility across multiple industries including automotive, healthcare, and robotics applications. Model cards can: 

• Teach students and help them understand real-world use cases 
• Inform policymakers and clarify intended use for non-model developers 
• Educate those interested in seeking the benefits of AI  

Mobilizing AI through model cards is a decisive and transparent step that companies can take toward the advancement of trustworthy AI.  

Improving and enhancing model cards   

We conducted market research to inform the improvements to existing model cards. While 90% of respondents in the developer sample agree that model cards are important and 70% would recommend them as-is, there is room for improvement to drive their adoption, use, and impact. 

Based on our research, existing model cards should be enhanced in two primary areas: accessibility and content quality. Model card users need model cards to be easily accessible and understandable. 

Accessibility 

Discovery is one element of model card accessibility that needs improvement. In releasing models, AI developers should be able to find and then promote model cards alongside their work. This is true of models introduced in research papers as well as models deployed for commercial use. 

Secondly, model cards need to be located where interested individuals can reference them. One of the ways NVIDIA promotes model cards is through the NGC Catalog. Models and model cards are located side-by-side in this same repository.   

Content quality

After a model card has been located, the next challenge for the user is understanding the information contained in it. This is particularly critical in the model evaluation stage before selection. Not understanding the information contained in the model card leads to the same outcome as not knowing that the information exists; either way, model users cannot make informed decisions. 

To address this, NVIDIA encourages using a consistent organizational structure, simple format, and clear language for model cards. Adding filterable and searchable fields is also recommended. When individuals can find the information contained in model cards, they are more likely to understand the software. According to our research, respondents liked and relied on the information contained in the model card when it was easily accessible and understandable.           

In fact, performance and licensing information were the two most important areas respondents wanted to see in model cards. Figure 2 shows how the StyleGAN3 model card devotes separate sections to performance and licensing.

After performance and licensing information, respondents felt that the section on ethical considerations was the most important category of information to include in model selection criteria. Within ethical considerations, respondents shared that they wanted more information about the datasets used to train and validate the model—particularly details regarding unwanted bias—as well as information about safety and security.   

Overview of Model Card++ 

Model Card++ is the improved NGC Catalog Model Card prototype that NVIDIA has developed over the last 9 months. In addition to the typical information given in the Overview section of a model card in the NGC Catalog, Model Card++ incorporates:

• The Plus Plus Promise (also known as the ++ Promise or the Triple P), describing the NVIDIA software development approach and the standards we hold ourselves to in all model development 
• Subsections detailing model-specific information concerning bias, explainability, privacy, safety, and security     

Figure 4 shows the ++ Promise, which will be embedded in every Model Card++.

The ++ Promise describes the steps that NVIDIA is taking to demonstrate the trustworthiness of our work embedded in design. The subcards outline: 

• Steps taken to mitigate unwanted bias
• Decision logic and example domains
• Provenance of training datasets and what type of data was collected and how 
• Development controls used and known restrictions     

This is not an exhaustive list but demonstrates the intent by design and commitment to standards and protections that value individuals, the data, and the NVIDIA contribution to AI. This applies to every model, across domains and use cases. 

Figure 5 shows an example of the Explainability subcard. Each Model Card++ will include a dedicated section of fields and responses for each of the subsections. What is shown in the response section for each field is not meant to represent a real-world model, but to illustrate what will be provided based on current understanding and the latest research.

The Explainability subcard gives information about example domains for an AI model, intended users, decision logic, and compliance review. NVIDIA model cards aim to present AI models using clear, consistent, and concise language.  

NVIDIA will start rolling out Model Card++ by the end of the year, with all commercial models using it by the end of 2023.   

How we built Model Card++

Model Card++ is the next generation of AI model card. It is the result of a disciplined, cross-functional approach in partnership with engineering, product, research, product security, and legal teams. Building on existing NGC model cards, we reviewed model cards from other organizations and templates, including GitHub, to find out what other information could be provided consistently. 

We worked with engineering to pilot what could be consistently provided in addition to information that is currently provided. We discovered that although our model cards have a section for ethical considerations, there is more that can be provided, like measures we took to mitigate unwanted bias. We also found we could describe dataset provenance and traceability, dataset storage, and quality validation.

We look to provide more details about dataset demographic make-up, performance metrics for different demographic groups, and specific mitigation efforts we have taken to address unwanted bias. We also worked with an algorithmic bias consultant to develop a process for assessing unwanted bias that is compliant with data privacy laws and coupled that with our latest market research.

As we built Model Card++, we also corroborated our work with market research by surveying developers who used our models and those from across the industry. We validated the desired information and structured it with our user design experience team to present it in a clear and organized format. We are excited about introducing Model Card++ to the world and hope to continue leading efforts that encourage inclusive AI for all.  

Get the latest updates about Model Card++ at the September GTC 2022 session, Ingredients of Trust: Moving towards Model Card++.

Text normalization (TN) converts text from written form into its verbalized form, and it is an essential preprocessing step before text-to-speech (TTS). TN…

Text normalization (TN) converts text from written form into its verbalized form, and it is an essential preprocessing step before text-to-speech (TTS). TN ensures that TTS can handle all input texts without skipping unknown symbols. For example, “$123” is converted to “one hundred and twenty-three dollars.”

Inverse text normalization (ITN) is a part of the automatic speech recognition (ASR) post-processing pipeline. ITN converts ASR model output into its written form to improve text readability. For example, the ITN module replaces “one hundred and twenty-three dollars” transcribed by an ASR model with “$123.”

ITN not only improves readability but also boosts the performance of downstream tasks such as neural machine translation or named entity recognition, as such tasks use written text during training.

TN and ITN tasks face several challenges:

• Labeled data is scarce and difficult to collect.
• There is a low tolerance for unrecoverable errors, as TN and ITN errors cascade down to subsequent models. TN and ITN errors that alter the input semantics are called unrecoverable.

TN and ITN systems support a wide variety of semiotic classes, that is, words or tokens where the spoken form differs from the written form, requiring normalization. Examples are dates, decimals, cardinals, measures, and so on.

Many state-of-the-art TN systems in production are still rule-based using weighted finite state transducers (WFST). WFSTs are a form of finite-state machines used to graph relations between regular languages (or regular expressions). For this post, they can be defined by two major properties:

• Mappings between accepted input and output expressions for text substitution
• Path weighting to direct graph traversal

In case of ambiguity, the path with the smallest sum of weights is chosen. In Figure 2, “twenty-three” is transduced to “23″ instead of “20 3.”

Currently, NVIDIA NeMo offers the following option for TN and ITN systems:

1. Context-independent WFST-based TN and ITN grammars
2. Context-aware WFST-based grammars + neural LM for TN
3. Audio-based TN for speech datasets creation
4. Neural TN and ITN

WFST-based grammar (systems 1, 2, and 3)

The NeMo Text Processing package is a Python framework that relies on the Python package Pynini to write and compile normalization grammars. For more information about the latest supported languages, see Language Support Matrix. For more information about how to extend or add your language grammar, see Grammar customization.

Pynini is a toolkit built on top of OpenFst, and it supports the export of the grammars into an OpenFST Archive File (FAR) (Figure 3). The FAR file can be used in a C++ production framework, which is based on Sparrowhawk.

Our initial version of TN/ITN system #1 does not take context into account, as that would make the rules significantly more complex, which requires extensive linguistic knowledge and deteriorates latency. If an input is ambiguous, for example, “1/4” in “The train leaves on 1/4” compared to “1/4 of a cup,” system #1 chooses the normalization deterministically without considering the context.

The system extends system #1 and incorporates context during normalization. The system outputs multiple normalization options in case of contextual ambiguity, which is rescored using a pretrained language model using Masked Language Model Scoring (Figure 4).

1. WFST generates all possible normalization forms and assigns weights to each option.
2. Pruning normalization options with weights higher than the threshold value “401.2″. In this example, we dropped “one/four”. It has a higher weight as it was not fully normalized.
3. LM rescoring picks the best among the remaining options.

This approach is similar to shallow fusion for ASR and combines the benefits of the rule-based and neural system. The WFST still limits unrecoverable errors while the neural language model resolves contextual ambiguity without the need for extensive rules or hard-to-get data. For more information, see Text normalization.

Table 1 compares the WFST+LM approach in terms of sentence accuracy with the previous system #1 (DetWFST) and a purely neural-based system (Duplex) on three datasets. Later in this post, we provide more details on system #4.

Overall, the WFST+LM model is most effective, particularly on EngConf, a self-collected dataset with ambiguous examples.

Figure 5 shows how susceptible the three methods are to errors. While the neural method is most affected by unrecoverable errors, such as hallucinations or omissions, WFST+LM is least affected by those and class ambiguity.

Audio-based TN (system 3)

Text normalization also comes in handy during the creation of new speech datasets. For instance, “six two seven” and “six twenty-seven” are both valid normalization options of ”627”. However, you must select the option that best reflects what is actually said in the corresponding audio. Audio-based text normalization provides such functionality (Figure 6).

Neural TN and ITN model (system 4)

One significant advantage of neural systems compared to rule-based systems is they are easy to scale if training data for a new language exists. Rule-based systems require much effort to create and may work slowly on some inputs due to combinatorial bursts.

As an alternative to the WFST solution, NeMo hosts a seq2seq Duplex model for TN/ITN and a tagger-based neural model for ITN.

Duplex TN and ITN

Duplex TN and ITN is a neural-based system that can do both TN and ITN. At a high level, the system consists of two components:

• DuplexTaggerModel:  A transformer-based tagger for identifying semiotic spans in the input (for example, spans about times, dates, or monetary amounts). [NEED LINK]
• DuplexDecoderModel: A transformer-based seq2seq model for decoding the semiotic spans into their appropriate forms (for example, spoken forms for TN and written forms for ITN).

The term duplex refers to the fact that this system can be trained to do both TN and ITN. However, you can also specifically train the system for only one of the tasks.

Thutmose tagger

The Duplex model is a sequence-to-sequence model. Unfortunately, such neural models are prone to hallucinations that could lead to unrecoverable errors.

The Thutmose Tagger model regards ITN as a tagging task and mitigates hallucination issues (Figures 7 and 8). Thutmose is a single-pass token classifier model that assigns a replacement fragment to every input token or marks it for deletion or copying without changes.

NeMo provides a method of dataset preparation, based on granular alignment of ITN examples. The model is trained on the Google Text Normalization dataset and achieves state-of-the-art sentence accuracy on both English and Russian test sets.

Tables 2 and 3 summarize evaluation results for two metrics:

• Sentence accuracy: An automatic metric that matches each prediction with multiple possible variants of the reference. All errors are divided into two groups: digit error and other error. Digit error occurs when at least one digit differs from the closest reference variant. Other error means a non-digit error is present in the prediction, for example, a punctuation or letter mismatch.
• Word error rate (WER): An automatic metric commonly used in ASR.

d-BERT stands for distilBERT.
Default is the default Google Text Normalization test set.
Hard is a test set with sampling of at least 1,000 examples for each semiotic class.

One-to-one correspondence between tags and input words improves the interpretability of the model’s predictions, simplifies debugging, and enables post-processing corrections. The model is simpler than sequence-to-sequence models and easier to optimize in production settings.

The sequence of input words is processed by the BERT-based token classifier, giving the output tag sequence. Simple deterministic post-processing gives the final output.

Conclusion

Text normalization and inverse text normalization are crucial for conversational systems and considerably affect users’ experience. This post introduced a novel way of handling TN task by combining the benefits of WFST and pretrained language models and a new neural tagging-based approach for tackling ITN task.

For more information, including code examples, tutorials, and documentation for the TN/ITN solutions discussed in this post, see the NVIDIA/NeMo GitHub repo.

Speaker diarization is the process of segmenting audio recordings by speaker labels and aims to answer the question “Who spoke when?”. It makes a clear…

Speaker diarization is the process of segmenting audio recordings by speaker labels and aims to answer the question “Who spoke when?”. It makes a clear distinction when it is compared with speech recognition.

Before you perform speaker diarization, you know “what is spoken” but you don’t know “who spoke it”. Therefore, speaker diarization is an essential feature for a speech recognition system that enriches the transcription with speaker labels. That is, conversational speech recordings can never be considered to be fully transcribed without a speaker diarization process because transcriptions without speaker labels cannot inform you who is speaking to whom.

Speaker diarization must produce accurate timestamps as speaker turns can be extremely short in conversational settings. We often use short back-channel words such as “yes”, “uh-huh,” or “oh.” These words are challenging for machines to transcribe and identify the speaker. 

While segmenting audio recordings in terms of speaker identity, speaker diarization requires fine-grained decisions on relatively short segments, ranging from a few tenths of a second to several seconds. Making accurate, fine-grained decisions on such short audio segments is challenging because it is less likely to capture reliable speaker traits.

In this post, we discuss how this problem can be addressed by introducing a new technique called the multi-scale approach and multiscale diarization decoder (MSDD) to handle multi-scale inputs.

Mechanism of multi-scale segmentation

Extracting long audio segments is desirable in terms of the quality of speaker characteristics. However, the length of audio segments also limits the granularity, which leads to a coarse unit length for speaker label decisions. Speaker diarization systems are challenged by a trade-off between temporal resolution and the fidelity of the speaker representation, as shown by the curve shown in Figure 2.

During the speaker feature extraction process in the speaker diarization pipeline, the temporal resolution is inevitably sacrificed by taking a long speech segment to obtain high-quality speaker representation vectors. In plain and simple language, if you try to be accurate on voice characteristics, then you have to look into a longer span of time.

At the same time, if you look into a longer span of time, you have to make a decision on a fairly long span of time. This leads to coarse decisions (temporal resolution is low). Think about the fact that even human listeners cannot accurately tell who is speaking if only half a second of recorded speech is given.

In most diarization systems, an audio segment length ranges from 1.5~3.0 seconds becausesuch numbers make a good compromise between the quality of speaker characteristics and temporal resolution. This type of segmentation method is known as a single-scale approach.

Even with an overlap technique, the single-scale segmentation limits the temporal resolution to 0.75~1.5 seconds, which leaves room for improvement in terms of temporal accuracy.

Having a coarse temporal resolution not only deteriorates the performance of diarization but also decreases speaker counting accuracy since short speech segments are not captured properly. More importantly, such coarse temporal resolution in the speaker timestamps makes the matching between the decoded ASR text and speaker diarization result more error-prone.  

To tackle the problem, we proposed a multi-scale approach, which is a way to cope with such a trade-off by extracting speaker features from multiple segment lengths and then combining the results from multiple scales. The multi-scale technique achieves state-of-the-art accuracy on the most popular speaker diarization benchmark datasets. It is already part of the open-source conversational AI toolkit NVIDIA NeMo.

Figure 2 shows the key technical solutions of multi-scale speaker diarization.

The multi-scale approach is fulfilled by employing multi-scale segmentation and extracting speaker embeddings from each scale. On the left side of Figure 2, four different scales in a multi-scale segmentation approach are performed.

During the segment affinity calculation process, all the information from the longest scale to the shortest scale is combined, yet a decision is made only for the shortest segment range. When combining the features from each scale, the weight of each scale largely affects the speaker diarization performance.

Multiscale diarization pipeline with neural models

Because scale weights largely determine the accuracy of the speaker diarization system, the scale weights should be set to have the maximized speaker diarization performance.

We came up with a novel multi-scale diarization system called multiscale diarization decoder (MSDD) that dynamically determines the importance of each scale at each time-step.

Speaker diarization systems rely on the speaker characteristics captured by audio feature vectors called speaker embeddings. The speaker embedding vectors are extracted by a neural model to generate a dense floating point number vector from a given audio signal.

MSDD takes the multiple speaker embedding vectors from multiple scales and then estimates desirable scale weights. Based on the estimated scale weights, speaker labels are generated. The proposed system weighs more on the large scale if the input signals are considered to have more accurate information on certain scales.

Figure 3 shows the data flow of the proposed multiscale speaker diarization system. Multi-scale segments are extracted from audio input, and corresponding speaker embedding vectors for multi-scale audio input are generated by using the speaker embedding extractor (TitaNet).

The extracted multi-scale embeddings are processed by clustering algorithm to provide an initializing clustering result to the MSDD module. The MSDD module uses cluster-average speaker embedding vectors to compare these with input speaker embedding sequences. The scale weights for each step are sestimated to weigh the importance of each scale.

Finally, the sequence model is trained to output speaker label probabilities for each speaker.

MSDD mechanism

In Figure 4, the 1-D filter captures the context from the input embeddings and cluster average embeddings.

In Figure 5, cosine similarity values from each speaker and each scale are weighted by the scale weights to form a weighted cosine similarity vector.

The neural network model MSDD is trained to take advantage of a multi-scale approach by dynamically calculating the weight of each scale. MSDD takes the initial clustering results and compares the extracted speaker embeddings with the cluster-average speaker representation vectors.

Most importantly, the weight of each scale at each time step is determined through a scale weighting mechanism where the scale weights are calculated from a 1-D convolutional neural networks (CNNs) applied to the multi-scale speaker embedding inputs and the cluster average embeddings (Figure 3).

The estimated scale weights are applied to cosine similarity values calculated for each speaker and each scale. Figure 5 shows the process of calculating the context vector by applying the estimated scale weights on cosine similarity calculated (Figure 4) between cluster-average speaker embedding and input speaker embeddings.

Finally, each context vector for each step is fed to a multi-layer LSTM model that generates per-speaker speaker existence probability. Figure 6 shows how speaker label sequences are estimated by the LSTM model and context vector input.

Figure 6, sequence modeling using LSTM takes the context vector input and generates speaker labels. The output of MSDD is the probability values of speaker existence at each timestep for two speakers.

The proposed speaker diarization system is designed to support the following features:

• Flexible number of speakers
• Overlap-aware diarization
• Pretrained speaker embedding model

Flexible number of speakers

MSDD employs pairwise inference to diarize conversation with arbitrary numbers of speakers. For example, if there are four speakers, six pairs are extracted, and inference results from MSDD are averaged to obtain results for each of the four speakers.

Overlap-aware diarization

MSDD independently estimates the probability of two speaker labels of two speakers at each step (Figure 6). This enables overlap detection where two speakers are speaking at the same time.

Pretrained speaker embedding model

MSDD is based on the pretrained embedding extractor (TitaNet) model. By using a pretrained speaker model, you can use the neural network weights learned from a relatively large amount of single-speaker speech data.

In addition, MSDD is designed to be optimized with a pretrained speaker to fine-tune the entire speaker diarization system on a domain-specific diarization dataset.

Experimental results and quantitative benefits

The proposed MSDD system has several quantitative benefits: superior temporal resolution and improved accuracy.

Superior temporal resolution

While the single-scale clustering diarizer shows the best performance at a 1.5-second segment length where the unit decision length is 0.75 seconds (half-overlap), the proposed multi-scale approach has a unit decision length of 0.25 seconds. The temporal resolution can be even more enhanced by using a shorter shift length that requires more steps and resources.

Figure 2 shows the concept of the multi-scale approach and the unit decision length of 0.5 seconds. Merely applying 0.5-second segment length to a single-scale diarizer significantly drops the diarization performance due to the degraded fidelity of speaker features.

Improved accuracy

Diarization error rate (DER) is calculated by comparing hypothesis timestamps and ground-truth timestamps. Figure 7 shows the quantified performance of the multi-scale diarization approach over the state-of-the-art and single-scale clustering methods.

The proposed MSDD approach can reduce DER up to 60% on two-speaker datasets when compared to the single-scale clustering diarizer. 

Conclusion

The proposed system has the following benefits:

• This is the first neural network architecture that applies a multi-scale weighting concept with sequence model (LSTM) based speaker label estimation.
• The weighing scheme is integrated in a single inference session and does not require fusion of multiple diarization results as in other speaker diarization systems.
• The proposed multi-scale diarization system enables overlap-aware diarization which cannot be achieved with traditional clustering-based diarization systems.
• Because the decoder is based on a clustering-based initialization, the diarization system can deal with a flexible number of speakers. This indicates that you can train the proposed model on two-speaker datasets and then use it for diarizing two or more speakers.
• While having all previously mentioned benefits, the proposed approach shows a superior diarization performance compared to the previously published results.

There are two future areas of research regarding the proposed system:

• We plan to implement a streaming version of the proposed system by implementing diarization decoder based on short-term window-based clustering.
• The end-to-end optimization from speaker embedding extractor to diarization decoder can be investigated to improve the speaker diarization performance.

For more information, see Multiscale Speaker Diarization with Dynamic Scale Weighting or see the Interspeech 2022 session.

Growing 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.