John Hughes

Accuracy Team Lead

Lawrence Atkins

Machine Learning Engineer

GPT-4 is a multimodal large language model from OpenAI that accepts both text and images as input. This technical deep-dive explains how GPT-4 works — covering its transformer architecture, RLHF training, and visual reasoning capabilities.

GPT4 - What is it?

GPT-4: How Does It Work? Architecture, Training & Capabilities Explained

Technical

Benedetta Cevoli

Senior Machine Learning Engineer

Our latest release, Ursa, breaks the accessibility barriers in speech technologies by offering ground-breaking accuracy for every voice. Irrespective of accent, dialect, and various demographic factors, Ursa is consistently the most accurate speech-to-text engine on the market with a relative lead of up to 30% compared to Amazon, Google, Microsoft, and OpenAI’s Whisper across the board.

Accessibility blog header image

Achieving Accessibility Through Incredible Accuracy with Ursa

Will Williams

Chief Technology Officer

We are excited to introduce Ursa, the world’s most accurate speech-to-text system. Ursa delivers unprecedented performance across a diverse range of voices. We observe relative accuracy gains of 22% and 25%* versus Microsoft and OpenAI’s Whisper respectively.

Ursa provides the world’s most accurate speech-to-text and is a breakthrough in accessibility, reducing the digital divide for voices that other speech-to-text systems struggle to recognize.

Ursa - header image

Introducing Ursa from Speechmatics

Headshot of Data Engineer, Steve Kingsley

Steve Kingsley

Data Engineer

Speechmatics' Data Engineer, Steve Kingsley, discusses Common Voice and OSCAR, and how their data can help to rapidly train languages.

How to Rapidly Train New Languages Using Common Voice and OSCAR

Speechmatics’ Data Science Engineer, Benedetta Cevoli looks at why speech technology still has some distance to travel in reducing inequality across different languages.  

Speech Technology in a Global World: Reducing Inequalities Across Languages

Accuracy Team Lead, John Hughes, explains how and why Word Error Rate as an accuracy measure is outdated and often misaligned with human judgment.

 In part one of a two-part piece, Accuracy Team Lead, John Hughes, explains how and why Word Error Rate as an accuracy measure is outdated and often misaligned with human judgment. Part Two looks at where we go next.

Problem WER image

The Problem with Word Error Rate (WER)

Caroline Dockes

Join Machine Learning Engineer, Caroline Dockes, as she explains the latest series of experiments aimed at moving speech recognition away from word-level vocabulary and towards subwords. 

Recognizing Rare Words: Experiments with Subword Units

Ellena Reid

Join Speechmatics Machine Learning Engineer, Ellena Reid, as she looks at the best way to turn a single node training setup into a robust, platform-agnostic, multinodal one. 

Nina Vanilla style 550x413px Webite Vanilla 550x413 PYTHON -F9FAFB

How to Successfully Achieve Multinode Training in PyTorch

Oliver Parish

  Machine Learning Engineer

Join Machine Learning Research Engineer, Oliver Parish, and Accuracy Team Lead, John Hughes as they look for ways to boost RNNT accuracy in speech-to-text using external Language Models.

How to Effectively Introduce External Language Models in the RNNT by Subtracting Internal Language Model Scores

Accuracy Team Lead, John Hughes, explains how we can move beyond Word Error Rate (WER) and looks for a better solution. 

Nina Vanilla style 550x413px Webite Vanilla 550x413 WER -F9FAFB

The Future of Word Error Rate

The Future of Word Error Rate (WER)

Ten questions to Stuart Wood, Director of Product, on the economics of speech at scale, how close local accuracy now gets to cloud, and who could benefit from Speechmatics On-Device.

On-device speech-to-text, in 10 questions

Sign In

Get started

Contact sales

Speech-to-Text API

Capabilities

Text-to-Speech API

Agent STT

Speech-to-text

Medical and healthcare-Navigation icon

Medical & Healthcare

Legal icon

Legal Transcription

technical-blog-icon

AI Voice Agents

media-and-captioning-icon

Media Distribution & Captioning

meeting-platforms

Note-Taking & Meeting Assistants

contact center icon

Contact Center Solutions

By Use Case

blog-icon

News & Articles

events-icon

Partnerships

contact-us-icon

Contact Us

careers-icon

Careers

about-speechmatics-icon

Find Out More About Speechmatics

case-studies-icon

Case Studies

Resources

Technical information

documentation-icon

service-status-icon

Service Status

Inclusivity

Community

Technical blog

announcement-icon

GitHub Academy

market-exposure-icon

Startup Program

Developers

Enterprise

Security

Pricing

Advanced speech recognition API with high accuracy across [<langcount>] languages.

Speechmatics Speech-to-Text

Get started with our speech recognition API

Free

Speechmatics

Advanced features for professional use ($0.24 per hour)

Custom solutions for large-scale deployments

We are Speechmatics. Global experts in deep learning and voice technology. Accurate speech-to-text for analysis and accessibility.

Speechmatics Social Sharing

Global experts in deep learning and voice technology. Accurate speech-to-text for analysis and accessibility.

Stay up-to-date with the very latest product news, company news and artificial intelligence (AI) research. Bookmark our technology blog today!