5 Reasons Deep Learning for Speech Recognition is Business-Ready Now

I'm frustrated. When I read from tech authors, advisors, and our competitor's blogs that End-to-End Deep Learning (E2EDL) Speech Recognition software is only being researched or not production-ready, I want to scream...

"Listen people! End-to-end deep learning speech recognition is ready and in production now, with customers running millions of hours of new audio transcribed per month."

Do these numbers make it sound like E2EDL is just a research project? Absolutely not. E2EDL has moved from research into stable production and shown the world that E2EDL is not just a pipe dream but a reality. In a previous post, we covered some of the technical differences between the traditional way of doing ASR-the one used by every company except Deepgram-and using E2EDL for speech recognition. But at this point, you might be saying, "Who cares, as long as the transcript I get is accurate?" If you only care about accuracy, I have good news for you-deep learning approaches to ASR are more accurate than traditional approaches.

But I'd guess you care about more than accuracy. You want a technology that can enable real-time communications. You want something that's cost-effective while also being easy to maintain and ready to adapt to future challenges. If that's true, I have even more good news-E2EDL approaches to speech recognition provide all of this and more. Let's dive in and talk about five of the key ways that deep learning for voice recognition can support your business.

Whether you're most interested in lower costs, higher accuracy, faster turnaround, easier scaling, or a future-ready technology, deep learning is the way to go.

E2EDL technology is much harder to develop initially but costs less to use. That's because DNNs can utilize hardware acceleration and GPUs to do multiple things at the same time, rather than running things in sequence like a CPU. Overall, you need less computing power than traditional ASR that runs on CPUs. This means you pay for less computer usage time to get transcripts back from your model or from a speech recognition API. Plus, you also save time and money on the model maintenance side, as you only have to maintain one thing, rather than a Franken-model composed of multiple parts.

For traditional ASR, "you get what you get." E2EDL allows you to maintain context through the entire process because you're not going through independent steps or models and hence the accuracy of each word and sentence improves. For example, deep learning is much quicker to train to focus on the speakers and transcribe the audio to get the important keywords correct.

That's because you only have to update a single model, rather than each step of a traditional ASR model. This makes it feasible to train a new model for specific use cases with very little effort, rather than having to tweak multiple, connected models to get the output you want. No other architecture can quickly train use case-specific models.

As mentioned above, E2EDL models are faster because it allows massive computing parallelization opportunities with GPUs compared to single threading on a CPU for the traditional ASR method. What does this mean for businesses? It means real-time transcription is possible, enabling conversation AI for use cases in call and contact centers. It also means that, even if you don't need low-latency transcription, transcripts of historical data can be turned around much more quickly than would be possible with traditional systems.

Because of the massive parallelization of GPU resources, E2EDL can be vertically and horizontally scaled more easily at a more cost-effective level. E2EDL can run 450 concurrent streaming transcriptions on just one T4 NVIDIA GPU, with only a nominal increase in latency. If you scale up a cloud service to process more data or use internal computing resources, you'll need to pay for a lot more computing power if you're using a traditional ASR system.

Most researchers agree that HMM-GMM has reached the limit of speed, accuracy, and overall improvement. HMM-DNN has some room for improvement left but must compromise speed, accuracy, or computing resources; i.e., you cannot get great accuracy at speed or high speed at a low computing resource cost. E2EDL, on the other hand, still has plenty of room to improve on accuracy, speed, and scale-up efficiency as we move into the future. E2EDL is tackling use cases that simply wouldn't be possible with older ways of doing ASR. For example, one customer is using us for transcriptions and IBM Watson for translations to create meeting translations and transcriptions, so everyone in a meeting can speak their own language while you can view the discussion in your language, in real-time! The speed and accuracy can only be achieved with E2EDL.

All of these features make deep learning the best speech recognition option available today for businesses of any size, from start-ups to enterprises. Production-ready E2EDL shouldn't be the best-kept secret out there. Discussions should be around how E2EDL can continually improve based on specific use cases and audio features, not on whether or not it's production-ready. In data science and machine learning, there's a truism that says you should go for the simplest algorithm or tool that gets you the results you need, even if it isn't the latest technology; sometimes, a simple linear regression model is more than enough. Deep learning ASR models are in the unique position of not only being the simplest option available-a single model that does everything, rather than a few different models strung together-but also being the most accurate and the most cost-effective. End-to-end deep learning for speech recognition is ready now! If you still don't believe me, you can try Deepgram out for free at console.deepgram.com or contact our STT experts if you want to explore training a custom model for difficult audio situations.