With so many nuances in speech, converting it into analyzable text is a huge technical challenge. How could Cedat 85 strengthen its position at the forefront of speech recognition technology?
Cedat 85 harnessed the power of accelerated servers from IBM to speed up the training of neural network models—enabling the company to take on more projects and provide results sooner.
2.4x fastertraining of models reduces time to market and boosts accuracy of solutions
Growsrevenue by increasing scope for new projects and experimentation
Billions of words/dayrecognized, driving novel applications for speech recognition
Business challenge story
Uncovering the secrets of speech
For over hundreds and thousands of years, humans have used speech to communicate complex ideas and emotions quickly and effectively. Speech alone, however, is very hard to analyze without first producing a written record, which takes time and resources.
Speech recognition technology aims to change this by automatically converting speech into written text that can be analyzed. For more than three decades, Cedat 85 has been making breakthroughs in speech-to-text solutions and helping clients in media, government, financial services, insurance, technology, marketing departments and contact centers to meet increasingly complex demands.
Enrico Giannotti, Managing Director at Cedat 85, explains: “Speech is a very sophisticated phenomenon. We speak using different languages, dialects and linguistic models. Because speech is so dynamic, with new phrases invented and existing words shifting their meaning every day, developing speech recognition solutions can be like hitting a moving target.”
Since 2002, Cedat 85 has built sophisticated neural networks to underpin its speech-to-text solutions, with its algorithms becoming more complex and more accurate over time. To take the next leap in development, the company wanted to speed up model training cycles. Previously, infrastructure limitations meant that it could take Cedat 85 weeks or even months to train a neural network model. With little to no insight into the model’s accuracy until the training was complete, this process was risky and time-consuming.
“We train our models using vast amounts of data,” comments Giannotti. “Even though we augmented our infrastructure with GPUs to speed up processing, it still took a long time to get results. Each project was so time- and resource-intensive that we had to postpone opportunities. With the growing popularity of analytics and voice-enabled solutions, new markets are emerging for speech-to-text all the time. To make sure we don’t miss out to competitors, we were determined to move faster.”
“ Using IBM Power Systems AC922 servers, we can train neural network models 2.4 times faster. And critically, these time savings don’t come at the cost of accuracy. ”
— Enrico Giannotti, Managing Director, Cedat 85
Getting word of the latest innovations
When Cedat 85 heard about IBM Power Systems AC922 servers, which are designed to support enterprise AI, it jumped at the chance to compare them to its existing GPU-accelerated infrastructure.
Giannotti recalls: “We were interested in finding out whether IBM could deliver on the promise of unprecedented performance for AI workloads with its latest generation of Power Systems servers. We tested the technology, and soon found that the answer was yes; they out-performed our previous environment significantly.”
Cedat 85 enlisted IBM Business Partner Computer Gross Italia to assist with the installation and, within a couple of days, the platform was up and running. Soon after, Cedat 85 won the tender for a landmark project: “Save our Sounds”, an initiative from the British Library to preserve the UK’s sound heritage. Cedat 85 was engaged to help establish a national radio archive, protecting the audio capital of past and future radio broadcasts from stations across the UK by transcribing the speech into text.
“Save our Sounds represents a big milestone for Cedat 85,” says Giannotti. “To deliver the project, we needed to enrich our models. For example, they had to recognize the rich variety of accents and dialects that exist in the UK and capture colloquial terms. The new, more powerful IBM infrastructure arrived at just the right time to support these efforts.”
“ Supported by IBM infrastructure, we produce billions of words of text from speech for our clients each day. By deploying the latest IBM Power Systems servers, we’re making sure that we continue to lead the way in development of speech-to-text solutions. ”
— Enrico Giannotti, Managing Director, Cedat 85
Speaking the language of AI
Cedat 85 is taking advantage of the data and compute-intensive IBM Power Systems servers to provide results to clients at great speed and quality. By reducing time to market for its solutions, the IBM infrastructure contributes to a vital competitive edge for the company.
“Using IBM Power Systems AC922 servers, we can train neural network models 2.4 times faster,” comments Giannotti. “And critically, these time savings don’t come at the cost of accuracy. We recently ran a pilot for a government client where we needed to tailor our models to accommodate different regional accents and dialects. Within a few weeks, we increased the accuracy of our English language model by almost 10 percent when applied to their use case, a process that would have taken us months in the past. This improvement is down to a combination of our AI engineers’ expertise, our proprietary algorithms and the super-fast processing of the IBM infrastructure.”
The shorter training times also reduce risk for Cedat 85, since engineers do not have to wait as long to find out if a neural network model needs fine-tuning. The company has also increased its capacity to take on projects, allowing it to capitalize on new opportunities.
“Deploying IBM Power Systems AC922 servers enables us to be more experimental, as we can try different things without worrying about tying up our infrastructure training a flawed model for months at a time,” explains Giannotti. “We can also be more proactive with our existing and prospective clients, helping us grow our revenues and giving us a return on our investment in IBM solutions within a few months.”
Looking to the future, Cedat 85 is investigating how it might integrate other IBM technology into its offerings. Specifically, the company is exploring how IBM Watson® tools might help it unlock additional value from text for customers.
Giannotti concludes: “Supported by IBM infrastructure, we produce billions of words of text from speech for our clients each day. By deploying the latest IBM Power Systems servers, we’re making sure that we continue to lead the way in development of speech-to-text solutions. Next up, we have the chance to team up with IBM to bring something even more special to the market.”
Cedat 85 is a market leader in speech-to-text technology and services based in Rome, Italy. The company has over 500 customers worldwide, 33 years’ experience and was designated one of the top 11 global players and one of the top five in Europe for Speech-to-Text Applications in Gartner’s 2018 competitive landscape. The company operates world-wide through its international brand Speech-i.
Take the next step
To learn more about IBM Power Systems, please contact your IBM representative or IBM Business Partner, or visit the following website: ibm.com/power
View more client stories or learn more about IBM Systems Hardware
To learn more about Computer Gross Italia SpA, visit: computergross.it