Service

Our service includes customized training for language specialists, adaptation of a language model to a new domain, optimization of the speech recognition systems, as well as development of new language components.


Workshops

Learning to talk is child’s play, mastering speech technology is not. It requires a combination of skills ranging from linguistics and computer science to engineering. EML offers special workshops for language specialists to train them for working with speech technology tools. The workshop participants learn to develop language models and acoustic models for a “new” language, to adapt language models to new applications, or to integrate speech technology into an application of their own.

For example, the Italian speech technology provider Cedat 85 used an EML workshop to develop a language model for Italian on the basis of the know-how acquired there and to integrate it into the existing transcription environment.

LINK: Cedat 85

 

Domain Adaption

An important component of speech processing is the language model. It includes the vocabulary and the probability of word sequences. This means that it covers all the words that can be recognized. It also contains an array of typical conversations in the target domain. Thus the language model describes WHAT is probably going to be said. The best recognition results are achieved when this information comes from real utterances recorded by the application.

EML offers their customers and partners to adapt the language model to the required application domains. All it takes is data, meaning texts from the respective application domain (words and sequences). Through domain adaptation, it is possible to achieve yet higher recognition rates. The customer can dynamically adjust his application domain using EML workplaces in many areas, for example in Speech Analytics.

Link: Products - EML Language Model Workplace

 

Recognition Optimization

A special service EML offers to its customers and partners is the recognition optimization using the acoustic model. The acoustic model depicts the sounds of a certain spoken language in form of statistics. It describes HOW people are going to speak. All modern transcription systems use hidden Markov model techniques for the acoustic model. These statistical models are based on a large volume of spoken data which has been transcribed. This includes speaker-typical data such as dialect, language style, gender, or age. Furthermore, the source of the acoustic signal is taken into account, e.g. whether it comes from a mobile phone, a noise cancelling microphone, or VoIP. And finally, ambient noise in the recordings (e.g. in an office, on a train, in a car, at a train station etc.) is also taken into account. This way, a precise and speaker-independent recognition can be guaranteed regardless of changing circumstances or ambient noise. The acoustic model is mostly application independent. However, adjusting it to a special application or acoustic environment can improve the recognition results.

Links: Products - EML Language Model Workplace, EML Transcription Workplace

 

Language Components


EML also develops language components for “new” languages to enhance the language portfolio. This includes acoustical models and language models. Furthermore, EML also develops language models for “new” languages for its customers. In most cases, the customer provides the audio data for his application. This way, the language model can be optimized for the relevant domain. The range of services offered by EML also encompasses customer-specific adaptation of a language model to a new domain using customer data.

Links: Products - EML Language Model Workplace, EML Transcription Workplace