Work Plan

The organization of the technical work is built around the main objectives of SpeDial. A cascaded waterfall model is used, namely, requirements specification in Task 4.1, design and implementation in WP2 and WP3, integration in Tasks 4.2 and 4.3, testing and evaluation in Tasks 4.4 and 4.5, and finally commercial exploitation and release in WP5. The work-packages and their tasks are built around the SpeDial main components, namely: IVR analytics (WP2) and speech services enhancement and customization (WP3).

The main objective of WP2 is to implement algorithms and tools for interactive voice platform analytics for call-center applications. The main components of the speech analytics platform are: emotion recognition, age/gender detection, call-flow/discourse analysis, talk-over analysis (barge-ins) and multilingual support. These core speech analytics technologies will be adapted and combined to identify hot-spots in the dialogue, i.e., areas where the user-system interaction breaks down or user satisfaction is low, as well as, identify the most probable root cause for the problem (grammar, prompt, dialogue flow). This information along with the key performance indicators of the dialogue system will be the input to WP3.

WP3 will provide algorithms and tools that use the output of the WP2 to enhance and customize the speech service, namely update prompts, grammars and call-flows in order to minimize dialogue hot-spots, maximize user satisfaction and reach target KPIs. Machine learning will be used to select prompts from a list, train/update statistical grammars from transcribed service data and optimize call-flow (asking most relevant questions first in the call-center application tree). In addition, user modeling/adaptation will be investigated, e.g., prompt and call-flow optimization for power users vs naive users. Multilingual speech services are also covered, where machine translation and crowd-sourcing is used to enhance prompts and grammars across languages. The algorithms and resources produced in WP2-3 are the main input of WP4.

In WP4 the IVR analytics and SDS enhancement modules are integrated into the platform. In addition, deployed speech services of the SME partners are enhanced and customized using the platform. The tools, algorithms and services are evaluated in terms of target KPIs and user satisfaction in Task 4.5. The two components of the platform (IVR analytics, service doctoring) are the main output of WP4 that are commercially exploited in WP5.

The generic Exploitation Model to be adopted in this project is shown next, with the required technological capacity decreasing from the inner to the outer circles. The exploitation model will additionally be guided by the Consortium Agreement, identified exploitation vectors, market analysis, consortium competencies and status of SpeDial outputs.

spedial exploitation model