New Deepspeech Training Scripts and Models for Welsh

Bangor University has just developed new training scripts and models that bring together the various features of DeepSpeech, along with CommonVoice data, and provides a complete solution for producing models and scorers for Welsh language speech recognition. They may be of interest to any other users of DeepSpeech that are working with a similarly lesser resourced language to Welsh.

The scripts:

  • are based on DeepSpeech 0.7.4
  • make use of DeepSpeech’s Dockerfiles (so setup and installation is easier).
  • train with CommonVoice data
  • utilize transfer learning
  • with some additional test sets and corpora, produce optimized scorers/language models for various applications
  • exports models with metadata

The initial README 4 describes how to get started.

We’d like to share also the models that are produced from these scripts which can be found at https://github.com/techiaith/docker-deepspeech-cy/releases/tag/20.06 4

At the moment these models are used in two prototype applications which the Welsh speaking community can install and try, namely a Windows/C# based transcriber and an Android/iOS voice assistant app 1 called Macsen. Source code for these applications using DeepSpeech can also be found on GitHub.

We are immensly grateful to Mozilla for creating the Common Voice and DeepSpeech projects.

Mozilla CommonVoice, Paldaruo ac Adnabod Lleferydd Cymraeg

Mae Mozilla, y cwmni rhyngwladol o Galifornia sy’n gyfrifol am y porwr gwe Firefox, newydd lansio eu cynllun CommonVoice amlieithog. Ar ôl cychwyn gyda Saesneg y llynedd, mae tair iaith newydd yn cael eu hychwanegu yn awr, sef y Gymraeg, Almaeneg, a Ffrangeg. Llwyddodd y Gymraeg i gyrraedd y brig oherwydd cymorth gan yr Uned Technolegau Iaith yng Nghanolfan Bedwyr, Prifysgol Bangor.

Rhagor o Leisiau i Common Voice
https://blog.mozilla.org/press-uk/2018/06/07/more-common-voices/#Cymraeg

Rydyn ni’n hynod o falch am CommonVoice Cymraeg ac yn awyddus iawn i gannoedd a miloedd o siaradwyr Cymraeg gyfrannu eu lleisiau drwy’r wefan neu’r ap.

 

Ond beth am Paldaruo? – ein ap torfoli sydd eisoes wedi casglu ers 2014 hyd at 38 awr o ddata lleferydd gan dros 500 unigolyn, ac sydd wedi helpu gwireddu meddalwedd cynorthwyydd personol digidol Cymraeg cod agored fel Macsen. Mae’r Uned wedi defnyddio gwaith Paldaruo i gynorthwyo Mozilla darparu CommonVoice ar gyfer y Gymraeg ac ieithoedd eraill llai eu hadnoddau eraill.

Un o’r heriau yw canfod a darparu testunau hwylus i’w ddarllen ond sy’n cynnwys ystod eang a chytbwys o ffonemau’r iaith. Ar gyfer y lansiad, mae 1200 promt gan yr Uned o fewn CommonVoice Cymraeg ond bydd angen mwy. Wrth i ni, a’r gymuned Cymraeg, gyfrannu rhagor o destunau a recordiadau i CommonVoice Cymraeg, rydyn ni’n rhagweld y bydd y corpws yn hwb sylweddol i weithgareddau ymchwil a datblygu adnabod lleferydd Cymraeg yr Uned ac eraill.

Y gobaith yw y bydd y bartneriaeth rhwng Mozilla a Phrifysgol Bangor yn tyfu, ac y bydd y gweithgaredd hwn hefyd yn symbylu cwmnïau mawr eraill i gynnwys y Gymraeg ac ieithoedd eraill llai eu hadnoddau yn eu cynlluniau rhyngwladol.

Cyfeiriad y wefan yw : https://voice.mozilla.org/cy ac mae’r ap ar gael o https://itunes.apple.com/us/app/project-common-voice-by-mozilla/id1240588326

 

Introducing Lleisiwr – Welsh Open Source Voice Banking and Text to Speech

In November 2017, The Language Technology Unit received a small grant from the Welsh Government’s Technology and the Welsh Language Fund, to work with the NHS as partners on a project to allow patients on the brink of losing their voice to bank their voice and then generate a personal digital synthetic voice. This had never before been availabe for Welsh speakers, and is a great step forward for Welsh speaking patients.

More information about this service can be found here including details for sofware developers about the package’s source code.

Here is a short video that shows you how to register for the service

There has been quite a favourable initial response on the social websites :

New Speech Resources

New speech resources have just been published by us under the Macsen project, funded by the Welsh Government. See details below. Enjoy!

HTK Acoustic Model

http://techiaith.cymru/htk/paldaruo-16kHz-2017-12-08.tar.gz

Lexicon

http://techiaith.cymru/htk/lexicon-2017-12-08.tar.gz

Prosodylab Aligner

There are also new HTK acoustic models included in the Welsh Prosodylab Aligner:

https://github.com/techiaith/Prosodylab-Aligner/tree/v2.0_paldaruo_4

Kaldi Acoustic Model

http://techiaith.cymru/kaldi/decoders/paldaruo_macsen/tri3-2017-12-18.tar.gz

Training code in GitHub

https://github.com/techiaith/kaldi-cy

RoboLlywydd

Or the ability to create your own naturally sounding Welsh language synthetic voices…

Ar part of our work on the Macsen project, we’ve created tools that will enable you to create naturally sounding Welsh language synthetic voices. The tools make it easy for you to prepare recordings scripts, record an individual’s voice, and with its knowledge of Welsh language pronunciation, build for you a Welsh language synthetic voice that sounds very similar to the recorded individual.

Here are examples of the voices of two members of the techiaith team having been synthesized with the new tools:

Male:

Female:

The team had the opportunity to demonstrate these tools at a recent SeneddLab 2017 event where a new voice was created within one hour, named ‘RoboLlywydd’ and used to speak the answers to questions about the National Assembly for Wales. Although the ‘RoboLlywydd’ name was just for fun, it showed that it’s possible to create and use many different individual voices within your own personal digital assistants. The following video talks more about this (especially after the fifth and a half minute):

We used an already open source system called MaryTTS which can now be used to create Welsh voices using the resources at the following GitHub repository:

techiaith/docker-marytts

Darlith Cymdeithas Wyddonol Gwynedd

Fe fydd Dewi Bryn Jones o Uned Technolegau Iaith, Canolfan Bedwyr, Prifysgol Bangor yn traddodi ar y pwnc;

Datblygu Adnabod Lleferydd ar gyfer y Gymraeg.

Mae’n gynyddol bosibl i chi siarad gyda dyfeisiadau fel eich ffôn neu gyfrifiadur er mwyn hwyluso defnyddio apiau, gwefannau a hefyd derbyn atebion deallus a pherthnasol i gwestiynau a ofynnwyd mewn iaith naturiol. Apple Siri, Microsoft Cortana, Amazon Alexa a Google Assistant yw rhai o’r cynnyrch a gwasanaethau masnachol poblogaidd sydd yn gyrru’r newid hwn gyda’r iaith Saesneg.

Yn y ddarlith hon bydd Dewi Bryn Jones o Uned Technolegau Iaith, Canolfan Bedwyr, Prifysgol Bangor yn cyflwyno’r gwaith sydd ym Mangor ar ddatblygu adnabod lleferydd ar gyfer cychwyn galluogi’r un ddarpariaeth i ddefnyddwyr Cymraeg. Swyddogaeth adnabod lleferydd yw trosi sain lleferydd unigolyn i destun ac felly bydd Dewi yn esbonio’r dulliau a’r data a ddefnyddir yn ogystal â chyflwyno’r canlyniadau diweddaraf.

Cynhelir y cyfarfod am 7.30 ar nos Lun Tachwedd 14eg yn ystafell 1.07 (llawr cyntaf), Canolfan Bedwyr, Y Ganolfan Reolaeth, Ffordd y Coleg, Bangor.

 

Introducing Macsen

During 2015-2016 we have been developing new resources that enable you to talking in Welsh with computers. See Start Speaking Welsh to your Computer, Towards a Welsh ‘Siri’

This is a technology which is becoming increasingly prevalent as the human voice is used more and more for question and answer systems on mobile phones and tablets, and voice control for such things as television sets, robots and dictation systems. If Welsh cannot be used in these environments, then the language will be excluded from the digital world and Welsh speakers will have no choice but to speak English with these devices.

In order to pave the way for new Welsh medium technologies we have produced a Welsh question and answer prototype, where a personal assistant called “Macsen” is able to answer questions such as what is the news or weather.

Here is a video that introduces Macsen and demonstrates it at work on a small Raspberry Pi computer:

All of Macsen’s code and resources are available on GitHub so that anyone can expand its capabilities and develop their own Macsen. The homepage for Macsen on the web and where you’ll know where to begin is:

http://techiaith.cymru/macsen

We will continue to work on speech recognition and other open resources for Macsen. Get in touch with us if you’re a software company, coding club, school or a hacker with an interest in including Macsen into your own software projects.

‘Macsen’ was developed within the  ‘Welsh Language Communications Infrastructure’ project which was funded by the Welsh Government and S4C.

Start speaking Welsh to your computer

We are developing Welsh language speech recognition as part of our Welsh Language Communications Infrastructure, sharing it here on the Welsh National Language Technologies Portal with other developers of Welsh language software and apps.

Today we are pleased to share the first version of a Welsh language speech recognition system

Julius Cymraeg (julius-cy)

This project is based on the Julius – an open source large vocabulary continuous speech recognition (LVCSR) system and the files, sripts required to its adaption for supporting to recognize Welsh language speech rather than English or Japanese.

mic_web
http://julius.osdn.jp/en_index.php

The first release allows julius-cy to recognize very simple questions and commands in Welsh concerning the weather, news, time, music as well as asking for a joke or a proverb. This means that julius-cy is limited to recognising specific sentences and vocabulary:

  • “BETH YDY’R TYWYDD HEDDIW?” ( “What’s today’s weather?” )
  • “BETH YW TYWYDD YFORY?” ( “What’s tomorrow’s weather?” )
  • “BETH YW’R NEWYDDION?” ( “What’s the news?” )
  • “FAINT O’R GLOCH YDY HI?” ( “What time is it?” )
  • “CHWARAEA GERDDORIAETH CYMRAEG” ( “Play Welsh music?” )

Future versions of julius-cy will attempt to support recognising dictation and more varied speech.

github_logo
https://github.com/techiaith/julius-cy

Everything you need to easily get started is available with very liberal licensing on GitHub.

Got to:

https://github.com/techiaith/julius-cy

 

This is amazing! How does it work?

The background page explains more about the internals of the first release:

https://github.com/techiaith/julius-cy/blob/master/CEFNDIR.md

You can try adding your own texts and questions for julius-cy to recognize after reading this!

Hmm. It doesn’t work very well for me. How can I help?

We are using very initial acoustic models in julius-cy, therefore it may be possible that julius-cy will not be able to fully recognize everyone’s speech successfully.

If this is the case, and you have not already contributed your voice to our Paldaruo Speech Corpus, then please use our Paldaruo ap (http://techiaith.bangor.ac.uk/paldaruo) on any iOS or Android device so that we can improve the acoustic models with your voice.

New Cloud based Welsh Machine Translation

coin-tinyPart of our Welsh Language Communications Infrastructure project is to improve the machine translation resources for leveraging some capabilities that are provided via English language based technologies.

As a result, the Welsh National Language Technologies Portal Moses-SMT machine translation’s capabilities are now available from the API Centre thus making it easy to integrate into your software including translation memory systems such as Trados, WordFast and CyfieithuCymru (TranslateWales)

Welsh<>English Moses-SMT joins a wide range of other language technologies API services such as Cysill (Welsh spelling and grammar checker), text-to-speech, parts of speech tagger, language detection, lemmatizer and Vocab to enhance Welsh support of your website, app and software.

api cloudSimilar to these other API services, you can get started by obtaining your API key (How to register for an API key) and follow the documentation and code examples we’ve prepared on GitHub. Please see: https://github.com/PorthTechnolegauIaith/moses-smt/blob/master/docs/APIArlein.md

Before you go ahead however, we’d like to emphasize once more the importance of quality control – It is your responsibility to ensure that this machine translation software is used appropriately, including the use of careful post-editing (see Quality Issues).

Demo

We have prepared a demo so that you can evaluate the machine translation engines.

Please see:  http://techiaith.cymru/translation/demo