Gaelic Algorithmic Research Group

Rannsachadh digiteach air a' Ghàidhlig ~ Goireasan digiteach airson nan Gàidheal

Author: levans3

The Acoustic Model and Scottish Gaelic Speech Recognition Results

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:

Audio Input Acoustic Model Output
Speaker says “She said hello” sh iy s eh d hh ah l ow

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:

ASR System Component(s) Output, given a speaker saying: “She said hello”
Acoustic Model Prediction sh iy s eh d hh ah l ow
+ Lexicon sh iy = she

s eh d = said 

hh ah l ow = hello

+ Language Model Prediction She said hello

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. 

Model

Training Corpus (hours of speech) Word Error Rate (WER) Accuracy

WER Reduction (from previous model)

A Clilstore (21) 35.8% 64.2%
B Clilstore 

+ Tobar an Dualchais (39.9)

31.0% 69.0% 4.8%
C Clilstore 

+ Tobar an Dualchais

+ Handwriting (63.5)

28.2% 71.8% 2.8%

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

Automatic Speech Recognition for Scottish Gaelic: Background and Update

By Lucy Evans

Since September 2020, a collaborative team from the University of Edinburgh (UoE), the University of the Highlands and Islands (UHI), and Quorate Technology, has been working towards building an Automatic Speech Recognition (ASR) system for Scottish Gaelic. This is a system that is able to automatically transcribe Gaelic speech into writing.

The applications for a Gaelic ASR system are vast, as demonstrated by those already in use for other languages, such as English. Examples of applications include voice assistants (Alexa, Siri), video subtitling, automatic transcription, and so on. Our goal for this project is to build a full working system for Gaelic in order to facilitate these types of use-cases. In the long term, for example, we hope to enable the automatic generation of transcripts and/or subtitles for pre-existing Gaelic recordings and videos. This would add value to these resources by rendering them searchable by word or topic. In this blog post, we describe our progress so far.

Data and Resources

There are 3 main components needed to construct a full ASR system. These comprise the lexicon, which maps words to their component phonemes (e.g. hello = hh ah l ow), the language model, which identifies likely sequences of words in the target language, and the acoustic model, which learns to recognise the component phonemes making up a segment of speech. The combination of these three components enables the ASR system to pick up on a sequence of phonemes in the input speech, map these phonemes to written words, and output a full predicted transcription of the recording.

  Input Output Prediction
Language Model The United States of <?> America
Acoustic Model Audio (Speaker says “Good Morning”) g uh d m ao r n ih ng

Of course, building these components requires resources. In terms of the lexicon, we are fortunate enough to have this resource already available to us. Am Faclair Beag is a digital Gaelic dictionary, developed by Michael Bauer, which includes phonetic transcriptions for over 30,000 Gaelic words. We simply pulled each word and pronunciation from this dictionary and combined them into a list to serve as our initial lexicon.

For training our language model (LM), we required a large corpus of Gaelic text. A LM counts occurrences of every 4-word sequence present in this text corpus, so as to learn which phrases are common in Gaelic. The following resources were drawn upon to build this:

  • The School of Scottish Studies Archives (UoE), which has provided hundreds of digitised manuscripts (via the earlier project, Building a Handwriting Recogniser for Scottish Gaelic)
  • The gd Corpus, which is a web-scraped text corpus assembled as part of the An Crúbadán project. This project aims to build corpora and other language technology resources for minority languages
  • Tobar an Dualchais/Kist o Riches, a collaborative project which aims to “preserve, digitise, catalogue and make available online several thousand hours of Gaelic and Scots recordings”. They supplied several hundred transcriptions of archive material from the School of Scottish Studies Archives

Finally, for training the acoustic model, we required a large number of speech recordings along with their corresponding transcriptions. This is so that the model can learn (with help from the lexicon) how the different speech sounds map to written words. We used recordings and transcriptions from the following sources to construct this dataset:

  • The School of Scottish Studies Archives (UoE) – see above
  • Clilstore, an educational website that provides Gaelic language videos at various different CEFR levels

A note on alignment

In order to train our ASR system to map speech sounds to written words, we must time-align each transcription to its corresponding recording. In other words, the transcriptions must be given time-stamps, specifying when each transcribed word occurs in the recording.

Time-aligning the transcriptions manually is lengthy and expensive, so we generally rely on automatic methods. In fact, we use a method very similar to speech recognition to generate these alignments. The issue here is that the automatic aligner also requires time-aligned speech data for training, which we don’t have for Gaelic.

We are fortunate in that we have been able to use a pre-built English speech aligner from Quorate Technology to carry out our Gaelic alignment task. As this was trained on English speech, it may be surprising that it is still effective for aligning our Gaelic data. However, despite noticeable high-level differences between the two languages (words, grammar etc.), the aligner is able to pick up on the lower-level features of speech (pitch, tone etc.), which are global across different languages. This means it can make a good guess at when specific words occur in each recording.

The alignment process – mapping text to audio.

Adapting the Lexicon

1. Mapping from IPA to the Aligner Phoneset

Because we are using a pre-built aligner on our speech data, we must ensure that the set of phones used to phonetically transcribe the words in our lexicon is the same as the set of phones recognised by the aligner’s acoustic model. Our lexicon, from Am Faclair Beag, uses a form of Gaelic-adapted IPA, whereas the Quorate aligner recognises a special, computer-readable set of English phones. For this reason, our first task was to map each phone in the lexicon’s phoneset to its equivalent (or closest) phone used in the aligner’s phoneset.

We first standardised the lexicon phoneset, mapping each specialised Gaelic IPA phone back to its standard IPA equivalent. We next mapped this standard IPA phoneset to ARPABET, an American-English phoneset that is widely used in language technology. This is the foundation of the aligner’s phoneset. We had to draw on our phonetic knowledge of Gaelic to create the mapping from IPA to ARPABET, because the set of phones used in English speech differs to that used in Gaelic: some Gaelic phones do not exist in English. For each additional Gaelic phone, we therefore selected the ARPABET phone that was deemed its ‘closest match’. Take the following Gaelic distinction between a non-aspirated, palatalised ( kʲ ) and non-aspirated non-palatalised ( k ) stop consonant, for example:

Gaelic IPA

(Gaelic phoneset)

Standard IPA

(global phoneset)

ARPABET

(English phoneset)

g k K
K

Our final mapping was from ARPABET to the aligner’s phoneset. Considering both of these phonesets are based on English, this was a fairly easy process; each ARPABET phone had an exact equivalent in the aligner phoneset. Once we had our final phoneset mapping, we converted all the phonetic transcriptions in the lexicon to their equivalent in the aligner’s phoneset, for example:

Word Original

(Gaelic IPA)

Standard IPA ARPABET Aligner
uisge ɯ ʃ gʲ ə ɯ ʃ kʲ ə UX SH K AX uh sh k ax
gorm g ɔ r ɔ m k ɔ ɾ ɔ m K AO DX AO M k ao r ao m

2. Adding new pronunciations

For our ASR system to learn to recognise the component phones of spoken words, we need to ensure that every word that appears in our training corpus is included in the lexicon.

Our initial phoneticised lexicon stood at an impressive 30,000 Gaelic words, however, the number of words in our training corpus exceeds 150,000. This leaves 120,000 missing pronunciations, many of which will simply be morphological variations on the dictionary entries. If our model were to come across any of these words in training, it would be unable to map the acoustics of that word to its component phoneme labels.

The ASR system maps the phones recognised by the acoustic model to words, using the pronunciations in the lexicon.

A solution to this is to train a Grapheme-to-Phoneme (G2P) model, which, given a written word as input, can predict a phonetic transcription for that word, based solely on the letters (graphemes) it contains. For example:

Input Output Prediction
h-uisgeanan hh uh sh k ih n aa n
galachan k aa el ax k aa n
fuaimeannan f uw ax iy m aa en aa n

We trained a G2P model using all the words and pronunciations already in our lexicon. The model learns typical patterns of Gaelic grapheme to phoneme mappings using these as examples. Our model achieved a symbol error rate of 3.82%, which equates to an impressive 96.18% accuracy. We subsequently used this model to predict the pronunciation for the 120,000 missing words, and added them to our lexicon.

Text Normalisation

1. Punctuation, Capitalisation, and other Junk

Our next tasks focused on normalising our text corpus. We want to ensure that any text we input to our language model is free from punctuation and capitalisation, so that the model does not distinguish between, for example, a capitalised and lowercase word (e.g. ‘Hello’ vs. ‘hello’), where the meaning of these tokens is actually the same. A simple Python programme was written for this purpose which, along with punctuation and capitalisation, also stripped out any junk, such as turn-taking indicators. Here is an example of the programme at work:

Input Output
A’ cur uèirichean ri pluga. a cur uèirichean ri pluga
An ann ro theth a bha e? an ann ro theth a bha e
EC―00:05: Dè bha ceàrr air, air obair a’ bhanca? dè bha ceàrr air air obair a bhanca

2. Digit Verbalisation

Another useful type of text normalisation is the verbalisation of digits. Put simply, this involves converting any digits in our corpus into words, for example, ‘42’ -> ‘forty-two’. An easy way of doing this is by using a Python tool called num2words. The tool is functional for verbalising digits into numerous languages, but unfortunately did not support Gaelic. For this reason, we coded our own Gaelic digit verbaliser, in order to verbalise the digits present in our text corpus. As the num2words projects welcomes contributions, we also hope to be able to contribute our code, so as to make the tool accessible to others.

Our digit verbaliser is currently functional for the numbers 0-100, and for the years 1100-2099. Also, as Gaelic uses both the decimal (10s) and vigesimal (20s) numbering systems, we ensured that our tool is able to verbalise each digit using either system, as specified by the user. We hope to eventually extend this to a wider range of numbers. The following examples show our digit verbaliser at work:

a) Numbers
Original Uill, tha, tha messages na seachdaine a chaidh agam ri phàigheadh agus bidh e timcheall air mu 80 pounds.
Vigesimal Uill, tha, tha messages na seachdaine a chaidh agam ri phàigheadh agus bidh e timcheall air mu ceithir fichead pounds.
Decimal Uill, tha, tha messages na seachdaine a chaidh agam ri phàigheadh agus bidh e timcheall air mu ochdad pounds.
b) Years
Original Bha, bha e ann am Poll a’ Charra ann an 1860.
Vigesimal Bha, bha e ann am Poll a’ Charra ann an ochd ceud deug, trì fichead.
Decimal Bha, bha e ann am Poll a’ Charra ann an ochd ceud deug ‘s a seasgad.

Current Work and Next Steps

After carrying out all the data and lexicon preparation, we were able to align our Gaelic speech data using Quorate’s English aligner. We have started using this to train our first acoustic models, and will soon be able to build our first full speech recognition system – keep an eye out for our next update!

Automatically subtitled video (using provided script)

However, aside from creating acoustic model training data, alignment can actually be useful for other purposes: it enables us to create video subtitles, for example. This kind of use case actually enables us to present our first observable results, which have been extremely encouraging. The videos in the link below exhibit our time-aligned subtitles, originally a simple transcription, separated from the video: click here to see examples of our work so far!

Powered by WordPress & Theme by Anders Norén

css.php

Report this page

To report inappropriate content on this page, please use the form below. Upon receiving your report, we will be in touch as per the Take Down Policy of the service.

Please note that personal data collected through this form is used and stored for the purposes of processing this report and communication with you.

If you are unable to report a concern about content via this form please contact the Service Owner.

Please enter an email address you wish to be contacted on. Please describe the unacceptable content in sufficient detail to allow us to locate it, and why you consider it to be unacceptable.
By submitting this report, you accept that it is accurate and that fraudulent or nuisance complaints may result in action by the University.

  Cancel