By Lucy Evans

In our last blog post, we outlined some of the data preparation that is necessary to train the acoustic model for our Scottish Gaelic speech recognition system. This includes normalization and alignment. Normalization is where speech transcriptions are stripped of punctuation, casing, and any unspoken text. Alignment is where each word in a transcription is stamped with a start and end time to show where it occurs in an audio recording.

After these steps, speech data can be used to train an acoustic model. Once combined with our lexicon and language model (as described in our last blog post), this forms the full speech recognition system. In this blog post, we explain the function of the acoustic model and outline two common forms. We also report on our most recent Gaelic speech recognition results.

The Acoustic Model

The acoustic model is the component of a speech recogniser that recognises short speech sounds. Given an audio input where a speaker says, “She said hello”, for example, the acoustic model will try to predict which phonemes make up that utterance:

The acoustic model is able to recognise speech sounds by relying on its component phoneme models. Each phoneme model provides information about the expected range of acoustic features for one particular phoneme in the target language. For example, the ‘sh’ model will capture the typical pitch, energy, or formant structure of the ‘sh’ phoneme. The acoustic model uses the knowledge from these models to recognise the phonemes in an input stream of speech, based on its acoustic features. Combining this prediction with the lexicon, as well as the prediction of the language model, the system can transcribe the input sentence:

Training the Acoustic Model

In order to train our acoustic model, we feed it a large quantity of recorded speech in the target language. These are split up into sequences of 10ms ‘chunks’, or frames. Alongside the recordings, we also feed in their corresponding time-aligned transcription:

Aligned Gaelic speech

Using the lexicon, the system maps each word in the transcript to its component phonemes. Then, according to the start and end times of that word, it can estimate which phoneme is being pronounced during each 10ms frame where the word is being spoken. By gathering acoustic information from every frame in which each particular phoneme is pronounced, the set of phoneme models can be generated. 

Training procedure for the Acoustic Model

Types of Acoustic Model: Gaussian Mixture Models vs Deep Neural Networks 

Early acoustic modelling approaches incorporated the Gaussian Mixture Model (GMM) for building phoneme models. This is a generative type of model, meaning that it recognises the phonemes in a spoken utterance by estimating, for every 10ms frame, how likely each phoneme model is to generate that frame. For each frame, the phoneme label of the model with the highest likelihood is output.

More recent, state-of-the-art approaches use the Deep Neural Network (DNN) model. This is a discriminative model. The model directly classifies each input frame of speech with a predicted phoneme label, based on the discriminatory properties of that frame (such as its pitch or formant structure). The outputs of the two models are therefore the same – a sequence of phoneme labels – but generated in different ways. 

The reason that the DNN has overtaken the GMM in speech recognition applications is largely due to its modelling power. DNNs are models with a number of different ‘layers’, and consequently a larger number of parameters. Parameters are variables contained within the model, whose values are estimated from the training data. Put simply, having more parameters enables DNNs to retain much more information about each phoneme than GMMs, and as such, they perform better on speech recognition tasks.

Another key difference between the two types of acoustic model is the training data they require. For GMMs, we can simply input recordings with their time-aligned transcriptions, as we already prepared using Quorate’s English aligner. On the other hand, training the DNN requires that every frame of each recording is classified with its corresponding Gaelic phoneme label. We obtain these labels by training a GMM acoustic model, which, once trained on the Gaelic recordings and time-aligned transcriptions, can be used for forced alignment. During forced alignment, each frame of the speech data is aligned to a ‘gold standard’ phoneme label. This output can then be used to train the DNN model directly.

Speech Recognition Results

Having carried out the training of our GMM and DNN acoustic models, we are now in a position to report our first speech recognition results. We initially trained our models using only the Clilstore data, which amounted to 21 hours of speech training data. Next, we added the Tobar an Dualchais data to our training set, which increased the size of the dataset to 39.9 hours of speech (NB: the texts in this data are transcriptions of traditional narrative from the School of Scottish Studies Archives, made by Tobar an Dualchais staff). Finally, we added data from the School of Scottish Studies Archives via the Automatic Handwriting Recognition Project to train our third, most recent model, on 63.5 hours of speech. 

We evaluated our models on a subset of the Clilstore data, which was excluded from the training data. This evaluation set comprises 54 minutes of speech, from 21 different speakers. Each recording was passed through the speech recogniser to produce a predicted transcription. We then measured the system’s performance using Word Error Rate (WER). The WER value is the proportion of words that the speech recogniser transcribes incorrectly for each input recording. The measure can also be inverted to reflect accuracy. 

As can be seen from the table below, our results have been encouraging, especially considering that DNN models perform best when trained on much larger quantities (100s of hours) of data. We are particularly pleased to report that our latest model passed below 30% WER (i.e. > 70% accuracy), an initial goal of our Gaelic speech recognition project. 

To showcase our speech recogniser’s current performance, we have put together some demo videos. These are subtitled with the speech recogniser’s predicted transcription for each video. Please note that the subtitles will have imperfections, given that we are using our speech recogniser (with 71.8% accuracy) to generate them. Take a look by clicking this link!

Demo video screenshot

Next Steps…

With just 2 months left of the project, the countdown is on! We plan to spend this time adding a final dataset to the model’s training data, with the hopes of further reducing the WER of our system. After this, we plan to experiment with speech recognition techniques, such as data augmentation, to maximise the performance of the system on the data we have collected thus far. Make sure to look out for further updates coming soon!

Acknowledgements

With thanks to Data-Driven Innovation Initiative for funding this part of the project within their ‘Building Back Better’ open funding call

A Linguistic Toolkit for Scottish Gaelic

Dr Loïc Boizou (Vytautas Magnus University) and Dr William Lamb (University of Edinburgh) have collaborated on a new bilingual website that provides a linguistic toolkit for Scottish Gaelic. Called Mion-sgrùdaiche Cànanachais na Gàidhlig or the Gaelic Linguistic Analyser, the site provides users with tools for analysing the words and structures of Gaelic sentences. The information provided by these tools can be used for additional natural language processing (NLP) tasks, or just for exploring the language further. This new website presents the tools together for the first time and provides users with two ways of interacting with them: a graphical interface and a command line method.

‘Like black magic’

The website’s development goes back to the late 1990s, when Lamb was working on his PhD. In order to investigate grammatical variation in Gaelic, Lamb constructed the first linguistically annotated corpus of Scottish Gaelic, spending over a year annotating 80,000 words of Gaelic by hand. He says, ‘It was a slog. Typing in 100,000 tags by hand… just don’t do it. I developed a nasty case of repetitive strain injury and vowed never to do this sort of thing by hand again.’ After returning to the University of Edinburgh in 2010, after 10 years at Lews Castle College Benbecula, he revisited his corpus to develop an automatic part-of-speech tagger and make the corpus available to other researchers. Today, the corpus is known as the ‘Annotated Reference Corpus of Scottish Gaelic’ or ARCOSG and is available freely online.

The corpus forms the backbone of two of the tools on the new website: the part-of-speech tagger and the syntactic parser. They were created using machine learning techniques, modelling the kinds of patterns that you find in Gaelic speech and writing. Lamb said, ‘what you can do today even with a relatively small amount of text is tremendously exciting. When we looked at developing a POS tagger in the 90s, we would have had to program each type of pattern manually to enable the computer to recognise it properly. Now, you can just run the corpus through a set of algorithms and the computer works the patterns out itself. It’s like black magic’.

Dr Will Lamb

The lemmatiser was developed in a different way, using a form of the popular online dictionary, Am Faclair Beag. Lamb explains: ‘When we were working on the part-of-speech tagger in 2013 or 14, Sammy Danso and I got in touch with Michael Bauer and Will Robertson, who put together the fantastic Am Faclair Beag. We were going to try to leverage some of the information in the dictionary, and they generously offered their data for this purpose. While that plan didn’t materialise, I was able to create a root finder or lemmatiser with it years later, which we used to help create the first neural network for Gaelic. The lemmatiser sat in the virtual cupboard for a while, until I was contacted by Loïc in 2017. Loïc wanted to create a proper Gaelic lemmatiser, and I was onboard.’

Dr Loïc Boizou

Dr Loïc Boizou is a Swiss French NLP specialist working in Lithuania (Vytautas Magnus University) who is interested in computational tools for under-resourced languages. He received his PhD in Natural Language Processing at Inalco (Institute of Eastern Languages and Civilisations) in Paris. About the project, he said, ‘I am very supportive of cultural diversity and Gaelic is one of the few endangered languages that provides serious opportunities for distance learning, thanks to Sabhal Mòr Ostaig. I really enjoyed learning the language and I decided to use my NLP skills to give it a bit of a boost. I learned about Will’s corpus and found we could cooperate very nicely.’

Roots, Trees and Tags

The website provides different ways of exploring  Gaelic text. Lemmatisation is simplest of the tools and involves retrieving a word’s root form. If you were to input a sentence like tha na coin mhòra ann (‘the big dogs are here’), the website would return ‘bi’, ‘cù’ and ‘mòr’ as the lemmas (root forms) of bha, coin and mhòra. The website also offers part-of-speech tagging, which provides grammatical information about words in a sentence. Using the previous example, the website’s algorithms would assign ‘POS tags’ to each word, as in the third tab-separated value in each line below (glossed in inverted commas):

tha	bi	V-p       'Verb: present tense'
na	na	Tdpm      'Article: pl masc def'
coin	cù	Ncpmn     'Noun: common pl masc nom'
mhòra	mòr	Aq-pmn    'Attributive adjective: plur masc nom'
ann	e	Pr3sm     'Prep pronoun: 3rd person sing masc'

The grammatical information in this example is quite precise, but such precision comes at a cost: the default tagger is subject to error about 9% of the time. For users who want simpler POS tags and more accurate tagging, the website also offers a ‘simplified tagset’ option, which provides 95% accuracy. The same sentence above, submitted with this option would provide the following:

tha	bi	Vp    'Verb: present tense'
na	na	Td    'Article: definite'
coin	cù	Nc    'Noun: common'
mhòra	mòr	Aq    'Adjective: attributive'
ann	e	Pr    'Prepositional pronoun'

In addition to lemmatisation and POS-tagging, the site also offers syntactic parsing, using a syntactically annotated corpus developed by Dr Colin Batchelor (Royal Society of Chemistry). Again, using the same sentence, the website returns the following if parsing is selected:

1	tha	bi	V-p	0	root
2	na	na	Tdpm	3	det
3	coin	cù	Ncpmn	1	nsubj
4	mhòra	mòr	Aq-pmn	3	amod
5	ann	e	Pr3sm	1	xcomp:pred

The number in the 4th column indicates which element in the sentence the word is governed by. In the case of tha, the number is 0, because it is the syntactic root. Both na and mhòra, on the other hand, are parts of a noun phrase governed by element 3, coin. This is a numerical way of displaying the kind of information that is often conveyed in a syntactic tree, such as in the example below. The information in column 5 indicates the function of the element in the sentence. For example, the function of coin is nsubj or ‘nominal subject’. More information on Dr Batchelor’s parser can be found here.

Syntactic tree for tha na coin mhòra ann

Next Steps

When asked what the next steps are for the language, Lamb explains that it’s an exciting time: ‘Well, this is really just an interim step and there is a lot to do. For a start, we hope to improve the accuracy of the tools gradually and perhaps augment them. Gaelic is, in some ways, in a very fortunate position when it comes to language technology. Advanced tools are starting to come online — like Google Translate, a handwriting recogniser and speech synthesiser — and we can exploit great resources like DASG, ARCOSG and recordings from the School of Scottish Studies Archives to push into territory that would have seemed like science fiction a few years ago.’

The dream is artificial general intelligence. ‘Elon Musk is famous for saying that one day, he’d like to die on Mars – just not on impact. Before I kick the proverbial bucket, I’d like to chat with a computer that has better Gaelic than I do’.

ARCOSG has been used for a range of projects including a voice synthesiser and syntactic parser. It has been newly revised and made compatible with the popular Natural Language Toolkit (NLTK): release available here.

A simplified version of the corpus has also been released, ARCOSG-S, which uses a less complex tag scheme (41 tags vs 246). It is available here.

With funding from UoE Challenge Investment Fund (Aug 2019), a small team of us have been busy developing the first handwriting recogniser for Scottish Gaelic. To do this, we have used Transkribus, a sophisticated, machine-learning based platform and on-line text repository.

Automatic transcription of Gaelic handwriting using Transkribus

The work began with the Digital Imaging Unit scanning about 2500 pages of handwritten manuscripts from the School of Scottish Studies Archives, supplemented by some additional scanning at the Centre for Research Collections.

Scanning manuscripts at the Centre for Research Collections

Once we received the texts, research assistant Michael Bauer manually transcribed about 18,000 words, which we used to generate our first Gaelic handwriting model. This achieved an impressive Character Error Rate (CER) of 2.53% – accuracy about 97.5%, but this was developed from and tested on one writer’s hand. We used this model to help transcribe a further 18,000 words and trained a second model. Again, this involved only one hand, but achieved a CER of 1.90%.

Using the updated model, we are moving towards our target of 500k words. We have focussed the transcription efforts recently on increasing the number of hands involved, so that our next model is more generalisable and useful. The project will finish in July 2020, when we intend to make the Gaelic handwriting recogniser available to the public through Transkribus.

Michael Bauer cataloguing the manuscripts

Project team

Dr William Lamb (PI): Celtic and Scottish Studies, LLC

Dr Beatrice Alex (Co-I): Edinburgh Futures Institute and LLC

Prof James Loxley (Co-I): English Literature, LLC

Dr Mark Sinclair (Consultant): Centre for Speech Technology Research (CSTR)

Mag Dr Muehlberger (Advisor): Transkribus (Innsbruck University)

Mr Michael Bauer (Research Assistant): Akerbeltz