Purpose

The purpose of these transcripts is to act as training data for speech recognition models, including our Welsh speech recognition models on Hugging Face. For that purpose, the transcriptions are more verbatim than traditional transcriptions or subtitles, so that the text reflects more closely what was actually said. This allows models to be evaluated against what was said rather than against a tidied-up version. A verbatim transcript can also be normalised into a cleaner style later, but what a clean transcript omits cannot be recovered.

The Process of Creating the Resource

The audio files were mainly collected from Welsh podcasts, after having gained the consent of the podcast owners and individual contributors to do so. We are extremely grateful to those people. In addition, some scripts were created which mimicked the pattern of news items and articles. These scripts were then read by Language Technologies Unit researchers in order to ensure that content of that type was included in the bank.

The audio files were run through our in-house automatic transcriber to segment the audio and create raw transcripts. Experienced transcribers then used both the ELAN 7.1 transcription software (available from https://archive.mpi.nl/tla/elan) and Trawsgrifiwr to listen to the recordings and correct the raw transcripts.

A Note About Content Anonymization

Out of respect to the contributors, we have anonymised all transcripts. It was decided to anonymize not only the names of individual people, but also any other Personally Identifiable Information (PII) including, but not limited to:

• Phone number
• Job titles/occupations
• Workplaces
• Names of public places
• Geographical location
• Dates/times

  When transcribing, all segments containing PII were marked with the <PII> tag, we then filtered out all segments containing a <PII> tag to ensure no personal information was published as part of this resource.

We have also randomized the order of the segments so that they are not published in the order they appeared in the original audio files.

Transcription Conventions

These transcription conventions were developed to ensure that the transcriptions were not only verbatim but also consistent. They were developed by referring to conventions used by the Unit in the past, conventions such as those used in the CorCenCC, Siarad, CIG1 and CIG2 corpora, and also through a process of ongoing development as the team undertook the task of transcription.

NOTE – as we have partially developed the conventions at the same time as undertaking the task of transcription the early transcriptions may not follow the latest principles faithfully. We intend to check the transcripts after we have refined the conventions.

Apostrophes

Apostrophes were not used to mark every single letter omitted by speakers. For example, gwitho (which is a pronunciation of gweithio) is correct, not gw’ith’o.

Rather, apostrophes were used to distinguish between different words that were otherwise spelled identically. For example we use an apostrophe in front of ‘ma (a pronunciation of yma) to distinguish it from ma’ (a pronunciation of mae), gor’o’ to distinguish between gorfod and the third person singular form of the present dependent tense gori, and pwysa’ to distinguish between the plural form of pwys and a number of possible verb forms of pwyso.

However, there is an exception to this rule, that being when spelling a word without an apostrophe would change the sound of the letter before or after the apostrophe, thus Cymra’g is correct, not Cymrag.

Tags

When transcribing, these tags were used to record elements that were external to the speech of the individuals:

• <anadlu>
• <anadlu i mewn yn sydyn>
• <aneglur>
• <cerddoriaeth>
• <chwerthin>
• <chwibanu>
• <chwythu allan>
• <clapio>
• <clirio gwddf>
• <crio>
• <cusanu>
• <ochneidio>
• <PII>
• <peswch>
• <sniffian>
• <tros-siarad>
• <twtian>

We anticipate that this list will grow as we transcribe more speech and as we come across more elements that are external to the speech of individuals.

Non-verbal sounds

Efforts were made to transcribe non-verbal sounds consistently. For example, yy was always used (rather than yrr, yr or err, or a mixture of those) to represent or reflect the sound made when a speaker was trying to think or paused in speaking.

The following were used in transcription:

• yy
• yym
• hmm
• m-hm

Again, we anticipate that this list will grow as we transcribe more speech and as we encounter more non-verbal sounds.

English words

We have surrounded each English word or phrase with asterisks, for example:

Dwi’n deall *sort of*.

Adapting English words as Welsh language infinitives

When speakers use English words as infinitives (by adding io at the end of the word for example) we have endeavoured to spell the word using Welsh spelling conventions rather than adding io to the English spelling of the word. For example we have transcribed heitio instead of hateio, and lyfio instead of loveio.

Correction of mis-pronunciations

To ensure that we adhere to the principles of verbatim transcription it was decided that we should not correct speakers’ mis-pronunciations. For example, in the following sentence:

enfawr fel y diffyg o fwyd yym efallu cam-drin

it is clear that efallai is the intended word, but it is transcribed as it is heard.

Punctuation

Full stops, question marks and exclamation marks were used when transcribing the speech.

We have surrounded all quoted words or phrases with ”, for example:

Dywedodd hi ”Dwi’n mynd” ond aeth hi ddim.

A note about our use of commas

As a comma is essentially a convention used for written text, commas were not used prolifically in transcription. Using a comma where one would expect to see it in a written text during transcription would not necessarily have reflected the individual’s speech. This should be borne in mind when reading the transcripts.

Individual letters

Letters were spelled out phonetically rather than being transcribed as capital letters.

That is, this is correct:

Roedd ganddo ow si di

not:

Roedd ganddo O C D

nor:

Roedd ganddo OCD

Numbers

Numbers were transcribed as words rather than digits, thus this is correct:

Y flwyddyn dwy fil ac ugain

rather than:

Y flwyddyn 2020

Half-finished words

Half-finished words are marked with a -. For example:

Ma’n rhaid i mi ca- cael diod.

Half-finished/restarted sentences

Half-finished sentences are marked with a .... For example:

Ma’n rhaid i mi ca’l… Ma’ rhaid i mi brynu diod.

Speaker interruptions

There are many examples of a speaker interrupting another speaker by using non-verbal sounds, words or phrases (such as m-hm, ie, ydi, yn union etc.) in the data. When the two speakers could be heard clearly and distinctly, a ... was placed at the end of the first part of the broken speech, and another ... at the beginning of the second part of the broken speech, as in the following example:

Ond y peth yw… M-hm. …mae’r ddau yn wir

When the two speakers could not be heard clearly and distinctly, the speech was omitted from the data.

NOTE: When the second part of a broken utterance appeared at the beginning of a segment, it was decided—for technical reasons—to either merge the segments or use ### instead of … at the start of the second part.

Specific terms

Although adhering to verbatim principles is the most important aspect of transcription, it is also important—as mentioned above—to aim for consistency, including in the spelling of certain word forms. It has therefore been decided to provide a list of word forms to be used during transcription, alongside forms that should be avoided. For example, it has been decided to transcribe swn rather than ‘swn (when the meaning is a contraction of byswn).

Swearwords

It should be noted that we have not omitted swearwords when transcribing.

The future

The bank is expanded regularly with new transcripts. We also intend to keep refining and harmonising the existing transcripts as the conventions develop.

Acknowledgements

We thank the contributors for their permission to use their speech. We are also grateful to the Welsh Government for funding this work as part of the Text, Speech and Translation Technology project for the Welsh Language.