Presentation is loading. Please wait.

Presentation is loading. Please wait.

This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement n° Reproducible.

Similar presentations


Presentation on theme: "This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement n° Reproducible."— Presentation transcript:

1 This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement n° 654237 Reproducible Automatic Speech Recognition workflows David Risinamhodzi – North west University – South Africa (David.Risinamhodzi@nwu.ac.za) e-Research Summer Hackfest – Catania (Italy)

2 Introduction & Overview 2 Automatic Speech Recognition training(ASR) Defn: Automatic speech recognition (ASR) can be defined as the independent, computer-driven transcription of spoken language into readable text in real time (Stuckless, 1994). Popular in the HLT community in SA Done for 11 official languages in SA Available speech corpus 50 GB high quality NCHLT speech data (16 kHz) 4.3 GB Lwazi telephony data (8 kHz)

3 Introduction & Overview (ASR) 3 Speech recognition requirements: Sufficient storage space for large audio and text datasets High Performance Computing (HPC): Many CPUs as most training and recognition is performed in parallel GPUs - use GPUs to train Deep neural nets (DNNs) High speed internet to move data around compute nodes Mechanism to manage datasets

4 Speech recognition tools : 4 HTK – Hidden Markhov Toolkit Bundled into “ASR_template” Libsvm – Support Vector Machine Delivered to endpoints with CODE-RADE

5 Scientific problem 5 Lack of collaboration Lack of exploitation of the available distributed computing Long hours of training systems on personal computers Research questions to be answered : Are speech recognition models reproducible ? How do speech recognition models vary according to different dictionaries and training ? Corollary issues : Provenance and publication of models ease of exploring ASR models – access for researchers Make access to national language resources easier

6 Computing and data model 6

7 TODO 7 ASR tools are working – but lots of integration for a researcher environment is necessary 1)Data ingestion into open-access repository 1) Dictionaries 2) Corpii 2)Web interface : 1) Select dictionary, corpus 2) Select task 3) Select specific parameters, or range 4) Task submission 3)Validation and comparison 4)Publication of the new analysis (model, accuracy, etc)

8 Implementation strategy 8 Web Interface to schedule training of jobs using specific datasets and recipes encompassing:- Ported application on the South African National Grid Boast of the fastest computer in Africa LENGAU Open Access Repository used to share data and results OneData platform is a good option to manage this Science Gateway used for job submission Kepler workflow management system may be a good option

9 Summary and conclusions 9 Goal is to encourage collaboration Fully exploit available resources and technologies Automate the process as much as possible Encourage use of Indigo cloud services in the near future.

10 Gracia! sci-gaia.eu info@sci-gaia.eu


Download ppt "This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement n° Reproducible."

Similar presentations


Ads by Google