Macsen is an open source Welsh language voice assistant. We are using this project to show what we can create when developing Welsh language speech technology and artificial intelligence. We publish relevant components and resources with open licences on the Welsh National Language Technology Portal, so that other developers can also use them. We are also undertaking further research to improve it, and to enable it in other environments. We have developed Macsen to work as an app on Android and iOS phones and tablets.
Macsen, the Welsh speaking personal assistant has 8 skills. It can
– Read the news
– Report on the weather
– Play Welsh language music on Spotify
– Set the alarm
– Tell you the time
– Give you the date
– Read the first couple of sentences of articles on the Welsh Wikipedia.
– Show a television program through the S4C Clic website.
It takes the news from the Golwg360 website, and you can ask it for the headlines, news from Wales, Britain or international news, and news about business, sport, health.
The weather comes from OpenWeatherMap.
The music comes from the Sptify website. AT the moment it recognises 21 bands and individuals, namely Alffa, Anhrefn, Anweledig, Bryn Fôn, Cadi Gwen, Gandelas, Cowbois Rhos Botwnnog, Cyrff, Gwibdaith Hen Frân, Gwilym Morus, Lleuwen, Mellt, Melys, Petrobas, Plant Duw, Sibrydion, Sŵnami, Y Bandana, Y Cyrff, Yr Ods, Yws Gwynedd. If you do not have a Premium Spotify account, it will sometimes play related music, rather than what you asked for – this is a feature of Spotify’s free accounts, and is not a speech recognition error.
You can also type as well as speak to ask questions or give commands to Macsen.
Macsen uses a number of different technologies to operate. It uses speech recognition to convert what you say into text. Then it uses intent parsing to recognise whether you asked for the news, weather, music or one of the other options. When Macsen needs to reply orally, it uses text-to-speech technology to speak the the appropriate response.
The app’s source code is available on GitHub as hopefully a useful resource for other developers:
Also , the source code for the app’s intent parser is available on GitHub:
We continue to improve the speech features, and if you are a Welsh speaker or learner, you can help us improve them by contributing recordings of your voice. You can do this within the app by clicking on ‘Hyfforddi’ there. This will lead you to read aloud the sentences recognised for the 7 skills within the app. We will use these recordings to create dev and test sets to train the speech recognition. If you want to contribute more than this, go to the Mozilla CommonVoice website (see below) to record sentences for the large recordings collection.
Further information on these technologies and the Welsh language may be found, in Welsh, in the Llawlyfr Technolegau Iaith (Handbook of Language Technologies) published by the Coleg Cymraeg Cenedlathol.
The app and speech development work was funded by the Welsh Government, and we thank them and the volunteers who have been contributing their voices to improve Welsh speech technology. Thanks are due also to Golwg360 and OpenWeatherMap for permission to use their on-line services.
Videos that explain the 7 skills (Welsh Language Only):-
1)Introduction to Macsen Application
2)Macsen’s Time Skill
3)Macsen’s Date Skill
4)Macsen’s Weather Skill
5)Macsen’s News Skill
6)Macsen’s Spotify Skill
7)Macsen’s Alarm Skill
8)Macsen’s Wikipedia Skill
HELP US DEVELOP MACSEN AND WELSH LANGUAGE SPEECH RECOGNITION
CONTRIBUTE YOUR VOICE THROUGH OUR CROWDSOURCING APP, ‘PALDARUO’ AND NOW ALSO THROUGH MOZILLA’S COMMON VOICE WEBSITE
The Paldaruo app and Mozilla’s Common Voice website (https://voice.mozilla.org/cy) are used to crowdsource recordings of various individuals speaking Welsh.
The Paldaruo app recordings are kept in the Paldaruo Speech Corpus. The corpus to date contains approx. 40 hours of recordings by nearly 570 individuals and has been used to train the speech recognition components within Macsen.
But we need more recordings, by more individuals, so that we can improve recognition accuracies and expand the range of texts and questions Macsen can recognize.
The Paldaruo app was developed two years before the start of Mozilla’s Common Voice project, and our first speech technology components were based on the Paldaruo corpus. We are very pleased to partner with Mozilla on Common Voice, which is an international project aiming to collect recordings for a number of different languages. We have transferred our original Paldaruo prompts to Common Voice, and you can choose whether to submit your voice through the Paldaruo app or through Common Voice.
Further instructions may be found on the Common Voice website. Here is a video showing how to use the Paldaruo app:
MACSEN ON THE RASPERRY PI
The first version of Macsen ran on the Raspberry Pi. This is demonstrated in the video below.
More information about Welsh language speech recognition resources are available from the following page on the Welsh National Language Technologies Portal.
All source code and other resources are available on GitHub so that anyone can join us and download, adapt and develop their own ‘Macsen’ system.
Let us know if you are, as a software company, a coding club, school or just as an enthusiast, using Macsen in your projects or activities.
Macsen Research Publications
BUILDING INTELLIGENT ASSISTANTS FOR SPEAKERS OF A LESSER-RESOURCED LANGUAGE, CCURL 2016 2nd Workshop on Collaboration and Computing for Under-Resourced Languages ‘Towards an Alliance for Digital Language Diversity’ LREC 2016, Portoroz, Slovenia. Paper
TOWARDS A WELSH LANGUAGE INTELLIGENT PERSONAL ASSISTANT: A brief study of APIs for spoken commands, question and answer systems and text to speech for the Welsh Government. 2015. Report