September 2022 Summaries

Speech-to-text technology has limitations that can be improved upon by considering various factors such as language bias, dialectal variations, and contextual nuances. To address these challenges, it's essential to approach languages with a flexible mindset and acknowledge the differences between languages, including their syntax, vocabulary, and evolution over time. By presenting models with diverse datasets and taking context into account, we can improve the accuracy of speech-to-text engines and deliver on our mission to understand every voice. Furthermore, being mindful of data bloat and its impact on training times and resources is crucial in optimizing our output while maximizing diversity.

It was with great sadness that Queen Elizabeth II's passing was announced, leaving a tremendous impact on many, especially her beloved family and the Commonwealth. The Queen gave her name to various scientific and educational projects worldwide, including the Queen Elizabeth II Graduate Scholarship in Science and Technology Program in Canada and The Queen's Award for innovation in technology in the UK. She witnessed immense technological change during her 70-year reign, honouring outstanding technologists such as Tim Berners-Lee and Sophie Wilson. As a female CEO, she was an inspiring figurehead who dedicated her life to service and respected globally for her devotion and leadership qualities.

IBC 2022 was a significant event in the broadcast media market, taking place after a three-year absence due to COVID-19. The global pandemic had a profound impact on businesses and investments, leading to fluctuations in the global markets and changing the landscape of the industry. This year's IBC show saw companies adapting to new working environments and networking to discuss industry changes, with many focusing on improving efficiencies and embracing technology solutions. Speechmatics showcased its new Language Identification feature at IBC, which identifies the predominant language in an audio file with a confidence rating. The event provided opportunities for content creators and technology providers to connect, launch partnerships, showcase new features, and explore the growing importance of AI in the broadcast world. Overall, IBC 2022 demonstrated the integration of speech recognition technology into mainstream broadcasting, marking a significant step forward for Speechmatics and the industry as a whole.

In its largest single language increase to date, Speechmatics has added 14 new languages to its accurate speech-to-text engine, empowering 340 million extra voices to use the technology in their native language. The company's expansion now supports over half of the world's population, outperforming Microsoft and Google's offerings. Speechmatics' new languages include lesser-spoken ones like Welsh and Basque, as well as constructed languages like Interlingua, aimed at preserving those communities and cultures for future generations. The company's Enhanced model has been trained to a higher standard than Google across shared languages, using self-supervised learning with little human input, resulting in an overall higher success rate in speech-to-text for all voices. Speechmatics plans to continue adding languages to its engine, aiming to reach 70% of the world's population in usability within three years.

Language Identification is a new feature that detects and labels the primary language from an audio file with a confidence rating using our Autonomous Speech Recognition (ASR). This feature builds on our existing Global-First language approach, which reduces the need to manually select the language for transcription, ensuring accurate transcription and avoiding inaccurate transcriptions due to transcribing files with the wrong language. By adding Language Identification, news becomes more accessible to everyone, contact centers become more efficient in extracting accurate data, and workflows are sped up by automatically identifying the predominant language in audio files rather than relying on human intervention. The feature supports 12 languages and provides a confidence score to show certainty of the predominant language.

As speech-to-text innovators, we are building systems and models that adapt to a broader range of voices to reduce the need for expensive, bias-creating, error-prone human intervention and labeled data. Speech recognition software cannot keep up with the times if it, like the languages it processes, does not evolve. To put this to the test, our Autonomous Speech Recognition (ASR) engine was compared to YouTube's automated captioning system across 24 videos, measuring its accuracy ratings against the video-streaming site's system. The results were conclusive, with our ASR achieving 90% or higher accuracy in all cases. As the world becomes increasingly globalized and online content grows, accurate speech-to-text software must continually update to keep up with evolving language and new words, making it a vital tool for accessibility and education. Our ASR features self-supervised learning, allowing it to learn from unlabeled data without human intervention, which gathers rich representations of speech from a wide breadth of voices and brings far-reaching accuracy. With the potential to save lives by providing accurate transcriptions in emergency situations, our software is poised to play a critical role in shaping the future of speech-to-text technology.