Terrill Dickey
November 1, 2024 05:30
Universal-2 addresses real-world needs and improves speech-to-text accuracy by focusing on structured data and important details over traditional word error rate metrics.
The introduction of Universal-2 represents a significant leap forward in speech-to-text technology and addresses real-world application requirements that go beyond traditional word error rate (WER) metrics. According to AssemblyAI, this advanced model targets the ongoing challenge of converting raw audio files into reliable, structured output.
Disadvantages of traditional indicators
Currently, the industry often claims voice recognition accuracy of over 90%. However, developers often encounter problems where the output, while technically correct, is not programmatically useful. For example, email addresses may be spelled as “Sarah dot Johnson at acme hyphenated core.com”, which can cause disruptions in data validation and program flow.
Universal-2 directly improves automation and user experience by focusing on providing output such as properly formatted emails and verified phone numbers from WER.
Evolution of speech recognition standards
The industry is obsessed with improving WER, but Universal-2’s slight improvement from 6.68% to 6.88% is far from the real effect. In a blind test, 73% of users preferred the output of Universal-2 and praised the application’s ability to deliver data in a format they could utilize immediately without further processing.
This model enables more sophisticated AI-based features by enabling applications to accurately distinguish between similar names and capture precise details such as timestamps.
Technological innovation driving Universal-2
Universal-2’s advancements stem from three key innovations:
- Tokenization for real speech: A new approach to handling repeated sequences improves the accuracy of phone numbers and product codes by up to 90%.
- Improved proper noun recognition: We double our supervised learning data and improve our neural architecture to better capture names and industry-specific terms.
- Neural network text formatting pipeline: Leverage a versatile tagging model and text range conversion model for improved punctuation, casing, and formatting accuracy.
Innovative business applications
The improvements in Universal-2 lead to real business benefits. In sales intelligence, models capture important details from customer interactions so you can accurately track and prioritize opportunities. Customer support benefits from accurate data capture, reducing the need for follow-up calls. In telehealth, this model minimizes administrative burden by ensuring appointments and prescriptions are recorded correctly.
Beyond the word error rate
Universal-2 is redefining what accuracy means in speech recognition by solving the last mile problem. This significantly improves the capture of proper nouns, alphanumeric characters, and formatting accuracy, allowing AI applications to go beyond WER and efficiently convert raw speech into structured business data.
Universal-2 can now be used to power the next generation of AI applications, giving developers the tools to build systems that not only record speech data in real time, but also understand and act on it.
Image source: Shutterstock