October 2023 Summaries

We recently added a new open test set to our evaluation suite which covers many languages and ran the numbers to compare ourselves against some others using publicly available data called the FLEURS dataset. We performed really well in underrepresented languages, outperforming our competitors by significant margins. We were also more accurate than Amazon, AssemblyAI, and Deepgram across all languages we offer. Our accuracy rate was 93.73% of the time when compared to every major ASR vendor. However, we recognize that these results are not entirely representative of real-world scenarios, as they are based on artificially clean data. To address this, we've made our test data more realistic by incorporating varied audio from different speakers, accents, and environments. Our tests show that Speechmatics is making significant improvements over OpenAI Whisper in these more challenging scenarios. We believe that accuracy is not the only metric to judge a provider, but rather usefulness in the real world. Our commitment is to achieve high accuracy regardless of input quality, even when it's low or noisy. To truly test this, we encourage users to try out our portal with their own audio and see the results for themselves.

At Speechmatics, they focus on building APIs that power various speech technologies, such as transcription, translation, and captions, for other companies' products. They enjoy working behind the scenes to provide specialist speech technology expertise. However, they also want to showcase their capabilities in new and innovative ways, which is where `the.Shed` comes in - a platform to share experimental prototypes and proofs of concept to spark ideas and imagination among their community. The company has created a few videos for `the.Shed`, including examples of how their API can be used to create agent assist tools and media monitoring dashboards. They invite users to contribute their own video ideas, showcasing their concepts in action, on Twitter.

Automatic speech recognition (ASR) can improve media companies' value far beyond acceptably accurate captioning when combined with artificial intelligence (AI) innovations. The growing preference for video content has led to a devaluation of audio, making captions essential for audiences who are accustomed to seeing them. AI media captioning has become accessible to smaller budgets, but many product teams still approach it as a cost-cutting measure rather than an opportunity to add value. Speech Intelligence combines ASR with capabilities powered by large language models (LLMs), enabling features like translation, summarization, and sentiment analysis. To unlock the full potential of Speech Intelligence, foundational accuracy is crucial, ensuring that ASR models can capture and understand different dialects, accents, and demographics accurately. By leveraging Speech Intelligence, media companies can deliver platforms that delight their partners, add real value, and stand out in a rapidly changing market.

The Evolution of Speech and the Unfulfilled Dreams of Sci-Fi Technology` explores the advancements in automated speech recognition, particularly by Speechmatics, which aims to harness the power of speech to transform various industries and aspects of human civilization. The technology has made significant progress in recent years, enabling seamless real-time speech recognition in multiple languages, summarization, sentiment analysis, topic detection, and more. However, despite these advancements, the technology still falls short of achieving the level of sophistication depicted in science fiction, such as C-3PO's fluency in over six million forms of communication or Samantha's conversational AI capabilities. The company believes that combining breakthroughs in speech recognition with large language models can make these science fiction dreams a reality, leading to a new era of artificial intelligence known as Speech Intelligence. This technology has the potential to revolutionize various industries and aspects of human civilization by empowering agents with useful information and key conversation playbooks, providing real-time monitoring of live streams, and comprehending topics being discussed in lectures. The ultimate goal is to create a seamless integration of multiple building blocks, including automatic translation, localization, inclusivity, expansion of geographic reach, and more, to unlock the value within speech data and enable interactions with computers to be as seamless as those with other people.

In a new era of AI, Large Language Models (LLMs) and Generative AI have reached mainstream, with applications in speech technology. The development of accurate transcription, comprehension, and transformation of speech data has been accelerated by LLMs, enabling the creation of cutting-edge products that harness the power of speech. Speech Intelligence is a new era for speech, where built with people in mind, this emergent technology will be ubiquitous and transform lives for the better. The technology aims to understand every voice, regardless of demographic, age, gender, accent, dialect, or location, and create active value with voice data. By combining accurate transcription with the power of LLMs and AI, Speech Intelligence moves from passively creating text from speech to actively taking action and creating value directly from voices. The technology represents a holistic view of taking audio from its source, transcribing, understanding, comprehending, transforming, and using it to build great products across industries. With vast use cases already being utilized today, Speech Intelligence has the potential to become one of the largest drivers of value within companies and teams.

The AI company, Speechmatics, aims to build a seamless AI stack called Speech Intelligence that connects the latest AI technologies to the spoken world, with a focus on improving speech recognition accuracy and providing a compounding advantage throughout the stack. The company has made significant progress in reducing Word Error Rates (WER) for its best-performing model, with a 50% reduction over the last two years. To achieve this, Speechmatics is investing in self-supervised learning to tackle the challenge of global transcription, which faces a unique obstacle due to the scarcity of labeled data, especially for diverse languages and speakers. The company believes that superb accuracy across any language will always be the preferred choice for conversational AI stacks, and it plans to pursue this goal with its next-generation self-supervised models and capabilities over one or more transcripts. Speechmatics also envisions a future where people can interact with technology seamlessly using their voices, without latency or misheard words, and is committed to investing in paradigm changes to make this vision a reality.

The contact center as a service (CCaaS) industry is on the cusp of significant change with the advent of Speech Intelligence, which combines artificial intelligence (AI) innovations with automatic speech recognition (ASR). This technology has the potential to revolutionize the sector by providing highly accurate transcription, interpretation, translation, automation, and pattern recognition capabilities. By leveraging ASR, CCaaS product teams can build valuable features such as AI routing, automated compliance requirements, analytics, and reporting, ultimately improving customer experience and agent working environment. The use of Speech Intelligence can also support innovative features like real-time transcription, keyword recognition, and language translation, making it a competitive differentiator in the market.

We've added an open test set to our evaluation suite, comparing ourselves against others using a publicly available dataset called FLEURS. We performed well in languages typically underrepresented, outperforming competitors by 8.45% on average and achieving higher accuracy than Amazon, AssemblyAI, and Deepgram in every language we support. However, we were surpassed by OpenAI Whisper in English by 0.48%. This led to a discussion about whether to share our test results, with some team members arguing that it could undermine our claims and damage our brand. Instead, we prioritized transparency and helping customers make informed decisions. We recognize the challenges of testing speech-to-text services, particularly with datasets that are too artificial and don't reflect real-world scenarios. To address this, we've made our test data more realistic by incorporating varied audio from different speakers, accents, and environments. Our results show that Speechmatics outperforms OpenAI Whisper in 8 out of 9 tests, making fewer errors than the competitor by an average of 32%. We believe that accuracy is not just about being the best at transcribing clean audio, but also about providing valuable transcripts regardless of input quality. Our goal is to create something useful and valuable, which requires testing our product with realistic data and acknowledging when we fall short. Ultimately, we encourage customers to try out our audio portal for themselves to see the results and make informed decisions.

Inclusivity in tech, particularly in Automatic Speech Recognition (ASR), has become increasingly important due to the risk of perpetuating discrimination through biased algorithms. ASR systems need to understand diverse voices and languages, including those with disabilities, to be truly beneficial. The current benchmarks for AI achievement do not adequately represent the perspectives of the global majority. Inclusivity in tech means that any piece of technology works independently of who is using it, regardless of language, accent, dialect, or surroundings. Speech recognition technologies have revolutionized our lives but can also be exclusionary, particularly for those with disabilities or from diverse backgrounds. To address these issues, self-supervised learning algorithms are being used to promote accessibility and reduce the reliance on labeled data dependency, which traditionally favors high-resource languages and demographics. Companies like Speechmatics are investing in this technology to create more inclusive ASR products that understand every voice, everywhere, regardless of language, accent, or surroundings.