Main menu

Automatic subtitling with speech recognition

Thanks to the evolution of speech recognition technology, there’s an easier and faster way of converting speech into text—such as for generating transcripts of audio content. Machines can perform this time-consuming tedious job accurately and quickly.

So who needs speech to text functionality?

Audio sections of libraries generate transcripts of faculty lectures and guest talks for students’ reference. Organizations like TED generate English transcripts of TED talks, which make it easier for people to translate the talks into other languages for a wider reach. Broadcasters generate closed captioning, also called subtitling for television programming, which lets the deaf and those who are hard of hearing follow the news and shows.

The beauty of transcripts is that they increase the potential audience of single language video content. Generating multilingual video subtitles and adding the option to display one of these per the consumer’s preference makes the content consumable by those who do not speak the language of the original audio stream. Incidentally, research shows that subtitling can ensure that more people watch a video to its completion.

Transcripts also make content consumable in mute mode. For instance, Oracle uses speech to text technology to generate closed captioning for proceedings of its enthusiastically followed technology conferences: Oracle Partner KickOff, JavaOne and Oracle OpenWorld. It acknowledges that subtitling gives “a global audience” access to the conference proceedings as well as “viewers tuning in at their desks requiring muted volume.”

Another plus point of transcripts is they enhance the value of content by making it searchable for certain words and phrases. Search engines like Google, Bing and even a company’s internal search engine can fully index transcripts, and thus make previously inaccessible audio content accessible. For instance, a civil engineering student might want to search faculty lectures for the key words “green building” or “glass house,” which is possible if the lectures have been transcribed.

Building speech to text functionality into apps could allow individual consumers to generate transcripts of live audio streams for their own consumption—such as helping attendees of multilingual conferences or participants of multilingual conference calls understand what’s going on.

GoVivace offers large vocabulary speech recognition technology that is especially well suited to generating accurate subtitling or audio transcripts, and thus speeding up what was once a labor intensive job. Accurately converting speech into text requires a speech engine that has a large and preferably, evolving vocabulary, so that it can understand practically any word that is fed to it. The GoVivace speech engine fits the bill.

Spare yourself from the listen-type-rewind cycle and raise the bar on subtitling in the process.


Voice-enable your mobile app to enable hands-free usage

Texting thumb is every smartphone user’s nightmare. All because mobile apps have invaded every facet of modern lives—socializing, shopping, banking, and so on. And, they promise to get still more versatile. Voice technology is predicted to power the next wave of mobile apps.

Consider for example a grocery store shopping mobile app. What every smartphone owner could use is a voice-enabled version, which would accept spoken input instead of compelling the user to fumble on a tiny keyboard. And the app would actually understand and respond to what has been said. Using an online grocery store’s app, they could just rattle off their shopping list and let the app search for the items and fill the cart. Then they could command the app to proceed to checkout, and voila! Without a single keystroke, the chore is done (leave aside waiting for the home delivery). No fumbling, no worn out thumbs. Nothing beats hands-free voice input—wouldn’t you agree?

Can an app really be so user-friendly?

Welcome to the talky online world populated by the likes of Google voice and siri, Apple’s personal assistant app. Technology to voice-enable mobile apps is already a reality.

GoVivace Inc. of McLean, VA has designed an advanced Automatic Speech Recognition engine that is specifically suited for voice-enabled mobile apps. Why’s that?

The key to building reliable and robust voice-enabled mobile apps is to construct a comprehensive application grammar and vocabulary, technical jargon for a set of pre-specified possibilities that the app will look up to understand the speech input. The more inclusive the grammar, the better the app will understand and seem intelligent to the user!

Say the user asks the voice-enabled mobile app of an online grocery store to add two packets of Oreo’s stuffed chocolate chip cookies to the shopping cart. A number of things happen behind the scenes. Just like siri or any other voice-enabled mobile app, the audio stream representing the spoken input is compressed and sent to a waiting farm of servers. Those servers are also notified the context in which the input was spoken. Putting together the context and the input, the servers quickly change their language model to suit the situation, and then convert the audio into text. The servers recognize that “two packets” is the quantity and “Oreo’s stuffed chocolate cookies” is the name of the item. Essentially, the item is looked up in the apps grammar and vocabulary representing the hundreds and even thousands of possible inputs the servers may have to process, and finally the cart is updated. It involves a lot of steps but everything happens so quickly that the app user doesn’t notice the number of steps involved, and happily goes on shopping.

The performance of voice-enabled mobile apps also depends on the quality of the speech recognition engine. Ideally, the engine must be capable of understanding natural language and adapting to variations in voice quality and spoken content. At the same time, it should be easy to use and integrate.

GoVivace’s Automatic Speech Recognition uses both grammars and a statistical language model to understand natural language, which helps build highly precise voice-enabled mobile apps.


Preventing financial fraud through voice biometrics

In the internet age, loss of personal credentials through security breaches (such as the recent Target breach) has become a common occurrence. This places individuals and financial institutions at an elevated risk of fraudulent activity. However some of this activity may be stopped in its tracks through voice biometrics. Consider that case when a user calls the wire transfer line to request a wire. The user presents some “secret” information, and then is able to complete the transaction. However under this setup the secret information could easily be stolen and then used against the account owner. However, if the bank deploys a voice biometric system, not only will the credentials of the user be stored in the system’s database, by also a small biometric key of the user that will be less than a kilobyte in size, shall contain the voiceprint of the user. Every time the user calls for a transaction, his voice shall be analyzed and a match will be created to tell if it is the same user. If an imposter makes the call, the system will clearly warn the agent to increase the security level, and use stricter authentication methods.

There are several features of modern voice biometric systems that make such a system not a dream of the future, but a real implementation that is in the process of happening. First, the voiceprints are small. It takes less than a kilobyte to store the voiceprint. It can be added to the same database that stores the passwords and other identifying information of the user, without much bloat in database size.

Second, the process of computing the match is extremely fast. A ten second of conversation at the start of the call may take just a fraction of a second to compare with the stored voiceprint, and tell if it is a likely match.

Third, there used to be a time when the same speaker will be easily rejected if they use a different handset to make the call or if the telephone channel is facing issues. However recent advances in voice biometrics have significantly reduced the impact of such factors, while maintaining the inter-speaker variability.

Financial institutions are already required to record conversations with their clients and store them for a certain period of time. If an institution determines to enhance their level of security through voice biometrics, these stored recordings could be used to enroll the voiceprints of all of their existing users. There will be no change to the user experience, though the call center agents will have an additional authentication metric available to them that allows the institution an additional layer of security without compromising convenience.

Subscribe to this RSS feed