Bangor Transcription Bank

This resource is a bank of 30 hours 20 minutes and 41 seconds of segments of natural speech from over 50 contributors in mp3 file format, together with corresponding 'verbatim' transcripts of the speech in .tsv file format. The majority of the speech is spontaneous, natural speech. We distribute this material under a CC0 open license.

Purpose

The purpose of these transcripts is to act as training data for speech recognition models, including our wav2vec models. For that purpose, transcriptions are more verbatim than what is seen in traditional transcriptions and than what is required for subtitling purposes, thus a bespoke set of conventions has been developed for the transcription work (see below ). Our wav2vec models use an auxiliary component, namely a 'language model', to further standardize the speech recognition model’s output in order that it be more similar to traditional transcriptions and subtitles.

We have provided 3 .tsv files, namely clips.tsv, train.tsv and test.tsv. clips.tsv contains all of our transcripts. train.tsv and test.tsv were created to provide 'standard' sets that allow users to compare models trained by different trainers fairly, i.e. they were created as a 'benchmark'. train.tsv contains 80% of our transcripts, and test.tsv contains the remaining 20%.

Here is an example of the data content:

audio_filename	audio_filesize	transcript	duration
f86a046fd0964e0386d8c1363907183d.mp3	898272	*post industrial* yym a gyda yy dwi'n ca'l deud	5092
f0c2310fdca34faaa83beca5fa7ed212.mp3	809720	sut i ymdopio felly, wedyn erbyn hyn mae o nôl yn y cartra	4590
3eec3feefe254c9790739c22dd63c089.mp3	1335392	Felly ma' hon hefyd yn ddogfen fydd yn trosglwyddo gyda'r plant bobol ifanc o un cam i'r llall ac hefyd erbyn hyn i'r coleg 'lly.	7570

There are four columns in the .tsv files. The first is the name of the audio file. The second is the size of the audio file. The transcript itself appears in the third column. The length of the audio clip appears in the last.

Here is the information about the columns.

Field	Explanation
`audio_filename`	The name of the audio file within the 'clips' folder
`audio_filesize`	The size of the file
`transcript`	Transcript
`duration`	Duration of the clip in milliseconds.

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 automated transcriber to segment the audio and create raw transcripts. Using Elan 6.4 (available from https://archive.mpi.nl/tla/elan), experienced transcribers listened to and corrected the raw transcript.

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.

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 asterixis, 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 expected 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

Individual letters were spelled out rather than being transcribed as individual 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.

Swearwords

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

The future

That this is an initial version of the transcript bank should be borne in mind when using this resource. We intend to refine and harmonize our transcripts further, and add yet more transcripts to the bank regularly over the next year.

Restrictions

In order to respect the contributors, by downloading this data you agree not to attempt to identify the speakers in the data.

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.

Datasets:

techiaith
/

banc-trawsgrifiadau-bangor