2023-04-04 POC Application SIA ASYA

Summary

SIA ASYA (asya.ai) is a Latvian company that specializes in the research, development, and commercialization of artificial intelligence (AI) models. The company's main product, pitchpatterns.com, is an AI-based automatic call analysis tool that is designed to improve sales and customer service. This tool is capable of analyzing sales and customer service calls, detecting the emotional tone of voice, biometric verification, and content, and automatically detecting problematic calls. Unlike other similar products in the market, pitchpatterns.com uses conversational dynamics analysis such as tone of voice, emotions, non-linguistic features like laughter, and does not rely on third-party services. The tool has already processed over 500,000 calls, and has been shown to improve conversion rates by 43% for clients. The company has attracted initial clients with 100+ licenses and 5000+ calls a day. Currently, the company uses direct sales and referrals as sales channels, but there is potential for scalable inbound sales. The company extensively uses Oracle Cloud Infrastructure (OCI) to deploy its services and train its models on GPU based compute nodes. The automatic backup functions and block storage that can be easily attached to any instance are particularly valued. With its award-winning scientists and UX designers and over 10 scientific publications, SIA ASYA is well-positioned to further develop and commercialize its unique AI technology for the multilingual European market. SIA ASYA possesses a range of proprietary AI models and datasets that are notable for their capabilities and efficiency. One significant aspect of their intellectual property is the absence of reliance on third-party services, which allows them to deploy their models on premises. With over 10 AI models and more than 4TB of datasets, ASYA.ai's intellectual property provides the best TSS for Baltic languages, speech enhancement, and emotion classification models, amongst other functions. Notably, asta.ai's speech enhancement models are superior to those of competitors such as krisp.ai, with the former delivering a PESQ score of 2.595, compared to the latter's score of 2.266. The company also has denoiser and speech enhancement models that enable them to process telephone calls with low-quality 8-16kHz audio, 32bit float. Furthermore, they have speech activity and VoiceID models, as well as end-to-end STT, which has a WER of 20% and CER of 10%, far better than Tilde's WER of 44% and CER of 25%. Additionally, ASYA.ai's proprietary punctuation and style model is capable of detecting and adapting to different writing styles. Overall, SIA ASYA intellectual property provides an impressive array of features that are suitable for improving sales and customer service, including emotions analysis, topic classification, and flagging models for suspicious calls. These unique advantages, combined with their ability to deploy on-premises and absence of reliance on third-party services, make them a highly competitive player in the market.

POC description

SIA ASYA relies on Oracle Cloud Infrastructure (OCI) to deploy its services and train its models. One of the advantages of using OCI is the automatic backup function and block storage, which can be easily attached to any instance. SIA ASYA uses GPU-based compute nodes to run its proprietary deep learning models on PyTorch. The company has been using older Tesla P100 nodes, which have 16GB of RAM, but is considering switching to newer Tesla A10 nodes, which have 24GB of video RAM and three times the number of cores. To compare the performance and cost-effectiveness of the two types of nodes, SIA ASYA plans to switch from seven Tesla P100 nodes to four Tesla A10 nodes over a two-month period, while also retaining its existing eight CPU nodes for web services. The NVIDIA P100 16GB has 3584 CUDA cores, while the NVIDIA A10 24GB has 10752 CUDA cores. Although the pricing of both cards is similar, the company is interested in understanding the potential cost savings and long-term costs of the services. As SIA ASYA has only had access to Tesla P100 nodes so far, the switch to Tesla A10 nodes represents a significant change in the company's cloud computing strategy. The evaluation period will allow the company to determine whether the newer nodes can deliver better performance at a lower cost, and whether they can meet the specific requirements of the company's deep learning models. By comparing the performance and cost-effectiveness of different nodes, SIA ASYA can make informed decisions about the best approach to cloud computing, with a view to optimizing performance and reducing costs over the long term. SIA Asya is requesting free of charge or discounted trial period of 2 months to configure long-term usage of OCI.

Plan

Week 1:

• Research and gather information, investigate the pricing and availability of the NVIDIA A10 GPU nodes

• Review the existing deep learning models and determine if they are compatible with the A10 nodes

• Create a test environment to benchmark the performance of the A10 nodes

• Start testing the performance of the A10 nodes with some sample deep learning models

Week 2:

• Deploy Denoising, Diarization models for API

• Evaluate the performance of the A10 nodes and compare it with the P100 nodes

• Document the results of the benchmarking tests

• Analyze the cost savings of using the A10 nodes over the P100 nodes

• Create a report outlining the performance and cost savings of using the A10 nodes

Week 3:

• Deploy Tone of voice Emotions, Speech Tempo, Speech Energy models for API

• Start testing the models on the A10 nodes and ensure that they are working correctly

• Monitor the performance of the A10 nodes and document any issues or challenges that arise

Week 4:

• Deploy Language detection models for API

• Optimize the models to take full advantage of the increased computing power of the A10 nodes

• Evaluate the performance of the optimized models and compare it with the original versions

• Create a report outlining the performance improvements achieved with the A10 nodes and the optimizations made to the models

Week 5:

• Deploy English and Russian STT models for API

• Continue to monitor the performance of the A10 nodes and identify any potential issues or bottlenecks

• Conduct stress tests on the A10 nodes to ensure that they are capable of handling high workloads

• Document the results of the stress tests and make any necessary adjustments to the configurations or models

Week 6:

• Deploy Latvian STT models for API

• Perform a final evaluation of the A10 nodes and compare the results with the original P100 nodes

• Determine the next steps, such as scaling up the infrastructure or continuing to use the A10 nodes for production

Week 7:

• Deploy Latvian and English Punctuation and Text style and text sentiment models for API

• If the results are inconclusive or not satisfactory, discuss potential solutions and next steps with the team to decide if further testing or a different approach is needed.

• If the results are positive, finalize the migration to Tesla A10 nodes and retire the Tesla P100 nodes.

Week 8:

• Deploy facial expression models for API

• Conduct a final review of the project to assess the success of the POC and identify any key takeaways or lessons learned.

• Prepare a report summarizing the project objectives, approach, results, and recommendations for future improvements or optimizations.

• Present the report to the team and stakeholders for feedback and discussion, and ensure that all project documentation is updated and organized for future reference.

Current usage