October 2022 Summaries

The Media Monitoring industry is rapidly growing due to the increasing levels of digitization and online media proliferation, with a projected 16.5% CAGR over the period 2022 to 2029. To succeed in this market, solution providers need to improve their offerings by addressing key priorities such as speed, global accessibility, avoiding missed mentions, and maximizing efficiency. Speechmatics offers fast real-time speech-to-text capabilities, supports multiple languages, enables custom dictionaries for accurate results, and streamlines batch transcription processes, aiming to help businesses differentiate their services and deliver on evolving customer expectations.

The NER model is an alternative to Word Error Rate (WER) for evaluating speech-to-text systems. It assigns a penalty level to errors based on their severity, with minor errors being easy to read through, standard errors disrupting the flow, and serious errors changing the meaning of text. The NER model requires human judgment to label errors, which can be subjective and difficult to automate. In contrast, WER follows a set of rules that can be fully automated. However, researchers have proposed alternative approaches using large language models (LLMs), such as few-shot learning and chain of thought reasoning, which can improve the accuracy of speech-to-text systems. The use of LLMs has led to significant advancements in natural language processing tasks, including reading comprehension and question answering. A new method uses a 2-shot prompt to automate the calculation of NER without human judgment, which may eventually replace WER for evaluating quality in Automatic Speech Recognition (ASR) systems.

As Speechmatics rounds out 2022, the company can be proud of its achievements, including multiple awards, $62m in Series B Funding, and the introduction of new features to its product offering. The addition of 14 new languages has brought the total to 48, providing language coverage for over half the world's population. This milestone is significant not only because of the number of languages but also due to the inclusionary nature of some of these choices, which reflects the company's commitment to understanding every voice. Speechmatics' approach prioritizes innovation and accuracy, with its Enhanced model being trained to a higher standard than Google across shared languages. The company aims to reach 70% global coverage by 2025 and will continue to release new languages only when they meet certain standards. With a forward-thinking mindset and a focus on inclusivity, Speechmatics is well-positioned for future success.

The media landscape is evolving rapidly with online content increasing, making captioning services crucial for differentiation and success in a highly competitive market. To stand out, providers must serve global audiences, improve viewer satisfaction through accurate speech-to-text, stay relevant by continuously learning language, and reduce risk by opting for best-in-class providers like Speechmatics. By doing so, businesses can tap into every market, deliver quality and efficiency, minimize reputational damage, and make the right move to differentiate their captioning services.

Whisper`, an open-source speech-to-text software launched by OpenAI, employs a unique approach that differs from current state-of-the-art systems. The model's training data contains 680,000 hours of audio, which is significantly larger than typical previous regimes. Whisper uses a simple autoregressive encoder-decoder structure and a cross-entropy loss function conditioned on the audio. This setup allows for impressive performance with supervised learning, but also raises questions about its generalization capabilities out-of-distribution. The authors suspect that a systemic failure in common evaluation protocols and overfitting to spurious correlations in the training data may be contributing factors to this issue. Despite these challenges, Whisper achieves human-level performance on certain tasks, demonstrating the potential of simple models with large-scale datasets. However, the plateauing nature of internet-scale supervised learning for English ASR systems suggests that more innovative approaches, such as self-supervised learning, are needed to overcome the limitations of current state-of-the-art systems.