Ursa 2: Elevating speech recognition across 50+ languages

Ursa 2 has landed

Speechmatics has released Ursa 2, our latest generation of speech recognition models, and with a major advancement in accuracy across all of our supported languages.

It represents a significant step forward from our previous generation, which themselves were industry leading across a diverse range of languages, dialects and accents:

• Ursa 2 achieved an 18% reduction in word error rate (WER) across 50 languages compared to our previous Ursa model.

• Our global Arabic language pack has undergone significant improvements, outperforming competitors across all dialects, including Egypt, Gulf, Levant and Modern Standard Arabic.

• We’ve added Irish and Maltese to our offerings, ensuring coverage of all 24 major European languages.

• We've reduced latency in our real-time transcription to <1 second, all while maintaining industry-leading accuracy.

Accuracy - the foundation for value

We have spoken before about the importance of accuracy, but it is worth restating.

Speech recognition and technology is often ‘foundational’ – applications, products and features are often built ‘on top’ of text generated from audio.

This means that errors in transcription can be spread within any system built on top of it, and the errors can compound. This is vital across industries, from media captioning and intelligence through to the newest generation of conversational AI agents and assistants.

Solid foundations, and highly accurate ASR are therefore vital to building valuable technology with speech as part of its stack. This is especially important for technology designed for industries that rely of technical terms and jargon, and where close enough is not good enough.

This solidity cannot be restricted to English spoken in a narrow range of accents – for us, consistency in our accuracy is key. We aim to provide speech technology that can be trusted, no matter the language, dialect, accent or noise environment.

How we approach the build of our models

The latest Ursa 2 models have continued our development of self-supervised learning (SSL) models.

This approach continues to give us huge benefits when faced with a bottleneck in the amount of labelled data available to us. In simple terms, there simply isn’t enough human-transcribed audio to cover every language, accent and dialect to guarantee high accuracy using supervised learning approaches to ASR.

Our SSL approach allows us to pre-train our models, by first immersing them in huge quantities of unlabelled data. Then, we couple this with more targeted training for a specific language, allowing us to reach high levels of accuracy, even if there is not thousands of hours of labelled data for that language.

A good analogy is that of a child learning a language.

When we are little, we do not receive formal language lessons (which here is the equivalent of received labelled data). Instead we simply are surrounded by people speaking and communicating with us and each other. Eventually, we spot patterns, imitate sounds ourselves and learn our first language like this.

Now imagine a child genius (think of Matilda as penned by Roald Dahl). Now, drop off Matilda at a large language school in a metropolitan area with hundreds of classrooms filled with students talking the world’s languages and return a few months later.

You can imagine that Matilda, despite not being enrolled in any formal classes, will have wandered the halls and by dropping into dozens of classes, over time will have not only picked up most of the languages spoken, but also have been able to spot patterns and commonalities between the languages, even if she hadn’t been taught this. You can well imagine that after this, Matilda would be able to learn new languages (even ones not taught at our fictional language school) incredibly quickly.

This is what we have done here at Speechmatics. This ‘pre-training’ gives us a huge head start with learning new languages, but also ensures we can achieve high accuracy across different languages and dialects.

For Ursa 2, we upped the amount of pre-training we have done, and increased the size of our ‘body model’.

We have combined this with other enhancements, including a complete shift over to GPU inference, and further reduction in our live transcription latency.

Universal improvements across every language

Assessing the accuracy of any speech-to-text provider can be more difficult than it first appears.

Though the FLEURS dataset is excellent in that it is publicly available and therefore transparent, it comes with its own challenges. Since it is publicly available, ASR providers can use it in the training of their models, and therefore accuracy scores can be artificially inflated compared to what you would see when using real data.

However, since it is often used as the benchmark, we will use it here to show improvements between Ursa and Ursa 2.

We have also used Word Error Rater (WER) as a measure. Accuracy can be calculated as 100 – WER.