Innovative contributions of the project

  • develop learning models for the Romanian language specific features, like the acoustic parameters, the probability structure of the n-grams and the probability structure of the morphological and the syntactical templates. Part-of-speech detection will represent another original fundamental contribution of this project.

 

  • apply techniques from signal processing in order to automatically partition and semantically annotate the speech signal.

 

  • adaptation to context. Our technology will be suited for a particular subset of the Romanian language. We aim at creating a system being able to work in noisy condition, be speaker independent and supply good transcription accuracy.

O2. Developing linguistic and acoustic resources for ASR and learning the artificial intelligence models

Within sub objective O2, low level modeling and training sessions will be performed for learning the acoustic parameters of the acoustic models, learning the probability structures of the n-grams, and of the morphological and syntactic templates, and accurate detection of the part-of-speech. When designing acoustic and linguistic models, experts on the acoustic and linguists will be employed. In the last year, an expert lawyer will join the project and will help designing experiments with the courtrooms and will contribute to extending the vocabulary for the judicial domain.

O1. Developing the automatic speech recognition technology

Within sub objective O1, the development team will first create the template of the technology, implementing a mock-up architecture with placeholders for various items that should be developed. Next, during the project, algorithms for the acoustic analysis, lexical analysis, morpho-syntactic analysis and efficient search will be developed and, when finished, placed in their holes in the core ASR technology.

About JustASR

While speech represents the most natural form of communication between humans, studying speech-based human-computer interaction presents a major interest for both academic research and commercial usage. Within this context, automatic processing of speech gains a major importance, being more and more embedded in various software products, in various domains from mobile phones, automated call centers, e-learning, medicine or justice. On the Romanian market, no software provider was able to deliver a speech-based product, even if the Ministry of Justice offered 2.5 Million Euros in a public tender held in 2010 for an automatic speech recognition (ASR) solution for courtrooms.  Within the IT strategy for 2013-2017 released in April 2013, the Ministry of Justice recalls the need for a software to automatically transcript the discussions inside the courtrooms.

This project aims to develop the core ASR technology for Romanian with applicability in the Romania’s courtrooms. We propose a speaker-independent solution for recognition of continuous speech, to enhance the automatic transcription of public hearing in courtrooms, as a measure of anti-corruption and public transparency.

The project will have the following two scientific objectives:

O1. Developing the automatic speech recognition technology

O2. Developing linguistic and acoustic resources for ASR and learning the artificial intelligence models