November 2022 Summaries

The text discusses the challenges and successes of implementing subword-level vocabulary in hybrid Automatic Speech Recognition (ASR) systems for English and German. The authors explore how using a word-level vocabulary is not feasible due to the constant evolution of language, and instead propose moving to a subword-level approach, which recognizes word pieces rather than entire words. They use Byte-Pair Encoding (BPE) as their tokenization algorithm and report promising results in German, where models trained with subwords perform well and recognize compound words that were previously outside the vocabulary. However, experiments in English show disappointing results, likely due to issues with long-range dependencies, word delimiters, and pronunciation, which need to be addressed in future research. Despite these challenges, the authors believe that this approach can have benefits for languages like German, where it has already shown promise.

The AI Bill of Rights is a non-regulatory blueprint that aims to guide the design, use, and deployment of automated systems to protect society from the most harmful aspects of AI, with five principles including Safe and Effective Systems, Algorithmic Discrimination Protections, Data Privacy, Notice and Explanations, and Human Alternatives. The guidelines are seen as a positive development for tech companies like Speechmatics, which already prioritizes data privacy and algorithmic fairness, and align with the company's own ethics and aims. The blueprint may be uniquely American, but other major regulators such as the EU are also moving forward with their own AI regulations, including the upcoming AI Act, which could have significant implications for the industry.

Gartner has provided insight into the future productivity and success of Contact Centers, highlighting the benefits of artificial intelligence (AI) in Contact Center as a Service (CCaaS) providers and partners. The report found that providers often combined natural language technologies to differentiate themselves, with a focus on detection, redirection, and summarization of calls. To provide the perfect service, a combination of factors is required, including proactivity, data analytics, and speech-to-text technology. Gartner emphasizes the importance of using data from customer interactions to improve efficiency and customer service. Independent Software Vendors can differentiate their solutions with better equipped speech-to-text supported by incredible AI, making it more secure for companies to serve their customers.

Multinode training, where multiple GPUs are used to train large neural networks, can be an effective way to speed up training time but requires careful implementation to avoid harming performance. To achieve this, companies must consider factors such as the number of nodes needed, networking setup, and containerization. Using InfiniBand with Remote Direct Memory Access (RDMA) can provide near-linear scaling across nodes, but requires careful software versioning and debugging. Additionally, errors can still occur during training, so it's essential to set up robust error handling mechanisms, such as webhooks, trap commands, and cleanup functions, to minimize downtime and increase the uptime of training runs.

While Unified Comms platforms have become a norm, the market is highly competitive, and providing a robust speech-to-text service that drives business value differentiation and accessibility for all is crucial. Many businesses are now seeking solutions with broad language and accent coverage, such as Speechmatics' technology available in 48 languages, to cater to their global workforce. By offering speed and accuracy, businesses can minimize manual human transcription costs and differentiate themselves against competitors. Additionally, features like Advanced Punctuation, Speaker Diarization, and Language Identification enable accurate notetaking and sales analytics, making it possible for businesses to include those who can't attend meetings or access data from archived audio files quickly and securely. By integrating Speechmatics' speech-to-text API, solution providers can push their Unified Comms offering to the next level and capitalize on the monumental growth of online and hybrid meetings.

Integrating external language models (LMs) into automatic speech recognition (ASR) systems can significantly improve accuracy while maintaining fast runtimes and low memory costs. Recent end-to-end models, such as the recurrent neural network transducer (RNNT), include an implicit internal language model (ILM) that is trained jointly with the rest of the network. However, when combining external LMs with RNNT models, it's essential to subtract the ILM scores to get the best accuracy results. This involves applying Bayes' rule and using scale factors λ₁ and λ₂ to balance the trade-off between acoustics and common word sequences. Approximating the ILM can be done through various methods, including removing acoustic data contributions or modeling it with an LSTM. Tuning the parameters λ₁ and λ₂ requires careful consideration of the dataset used, as they can significantly impact performance. By carefully balancing the contributions of external LMs and ILMs, ASR systems can achieve significant improvements in accuracy while maintaining fast runtimes.

Speechmatics' best-in-class speech recognition technology has improved significantly, enabling accurate speech-to-text for children's voices, reducing costs and time wasted on reviewing and editing captions. The company has expanded its language support to 48 languages, including 14 new additions, making it a global solution for educators and students worldwide. By choosing Speechmatics, EdTech product leaders can deliver a differentiated user experience that addresses the largest market possible, empowering educators and students with the right tools. The company's speech-to-text technology has been praised by top Video Distribution Platforms, including Udemy, which recently extended its use to Spanish, Portuguese, French, German, Italian, and Japanese courses. With Speechmatics' 6 simple steps for integration, product leaders can easily set up their solutions and differentiate themselves in the market.

This November, Speechmatics has released several updates to its language packs, expanding its coverage to 48 languages, including 14 new additions such as Bashkir, Basque, and Mongolian. The company aims to reach 70% usability in the next three years. Improvements have also been made to 20 existing language packs, increasing accuracy levels and adding new features like improved formatting of numeric entities. A Real-Time SaaS offering has been launched, providing fast and accurate speech-to-text results deployed within a secure public cloud environment. Additionally, the Batch SaaS has seen significant speed improvements, reducing turnaround time by up to 75%.

Speechmatics has launched its Real-Time SaaS API, enabling customers to incorporate real-time transcription into their operations using the cloud, greatly lowering barriers to entry for smaller businesses and making cloud-based speech-to-text more accessible and inclusive. The new API offers accurate real-time transcription without compromising on quality, providing a sliding scale for balancing speed with accuracy to meet specific needs. This rollout enables deployments of any size, from individual streams to major enterprises, and is additive to the already available Batch API in the cloud. With its self-service portal, customers can easily demo or integrate the API using their preferred programming language, including Python. The launch aims to accelerate customer cloud transformations while providing high-quality speech-to-text services to a wider range of users.

Our company has introduced a new Real-Time SaaS speech-to-text offering that provides fast and accurate results in a public cloud environment, opening up our technology to businesses of all sizes. Our on-premises offering had consistently outperformed Google and Microsoft for years but required infrastructure to achieve high results. The new SaaS approach simplifies adoption and allows organizations to access the most accurate and fast speech-to-text currently available. With latency as low as a second, users can see words in real-time without sacrificing accuracy. Our Real-Time SaaS provides instant rewards with actionable data and a flexible API that supports 48 languages, including less commonly recognized dialects. The solution is secure, easy to use, and accessible to all sizes of organizations, with a free trial available for those interested in trying it out.