Key words can help us remember important points in a transcript. But can AI find them?

With all the discussions around ChatGPT and Automatic Speech Recognition (ASR) it would seem that Large Language Models (LLMs) as part of a collection of Artificial Intelligence (AI) models could provide us with language understanding capability.   Some models appear to be able to answer complex questions based on the amazing amounts of data or information collected from us all.   We can speak or type a question into Skype and Bing using ChatGPT and the system will provide us with an uncannily appropriate answer.   If we want to speed read through an article the process of finding key words that are meant to represent a main point can also be automated as can a summarisation of content as described in the ‘ChatGPT for teachers: summarize a YouTube transcript’    

But can automated processes pick out the words that might help us to remember important points in a transcript, as they are designed to do when manually chosen[1]?

We rarely know where the ASR data comes from and in the case of transcribed academic lectures, the models used tend to be made up of large generic datasets, rather than customised educational based data collections.   So, what if the original information the model is gathering is incorrect and the transcript has errors or what if we do not really know what we are looking for when it comes to the main points or we develop our own set of criteria and the automatic key word process cannot support these ideas.   

Perhaps it is safe to say that keywords tend to be important elements within paragraphs of text or conversations that could give us clues as to the main theme of an article.  In an automatic process they may be missed where there is a two-word synonym such as a ‘bride-to-be’ or ‘future wife’ rather than ‘fiancée’ or a paraphrase or summary that can change the meaning:

Paraphrase: A giraffe can eat up to 75 pounds of Acacia leaves and hay every day.[2]

Original: Giraffes like Acacia leaves and hay and they can consume 75 pounds of food a day.

Keywords are usually names, locations, facts and figures and these can be pronounced in many different ways when spoken by a variety of English speakers[3].  If the system is using a process of randomly learning from its own large language model there maybe few variations in the accents and dialects and perhaps no account of aging voices or cultural settings.  These biases have the potential to add yet more errors, which in turn affect the relevance of chosen key words generated through probability models.    

Teasing out how to improve the output from AI generated key words is not easy due to the many variables involved.   We have already looked at a series of practical metrics in our previous blog and now we have delved into some of the other technological and human aspects that perhaps could help us to understand why automatic key wording is a challenge.    A text version of the mind map below is available.

Figure 1. Mind map of key word issues including human aspects, when working with ASR transcripts. Text version available




Expanding the metrics for evaluating ASR recordings to highlight errors possibly related to bias in training data.

In the previous blog we discussed some of the metrics that we felt needed to rbe explored in order to carry out a fuller evaluation of ASR recordings in order to try to address some of the issues occurring in the output to captions and transcriptions.

Recently we have developed a range of practical metrics evaluated by a series of scores and value-added comments. This was felt necessary to solve the issue of selection bias in ASR that seems to highlight errors due to pronunciation differences affected by age, gender, disability, accents and English as a foreign language when listening to lecturers across a range of subjects. It is hoped these can be addressed by providers using differently biased input data that is customized, instead of using one single accuracy percentage to denote the performance of the ASR services.  Evaluators also need to be aware of these issues and suggest the need for more inclusive training data to enable corrections to automatically occur in a proactive manner.

In the table below the list of items that may be used in a review has been expanded well beyond those usually used to find the type of word errors, omissions or additions that are occurring.

Speaker SpeechEnvironment NoiseContent – What is expressedTechnology              HardwareRecording
Pronunciation Clarity Speed Loudness Pitch Intonation Inflection Accent,
Age, Gender, Use of Technology Too far away / near the microphone   
Ambient noise/continuous Reverberation Sudden noise Online/Offline User device Room system Conversation vs Presentation Single speaker Overlapping speakers Multi-speakersComplexity Unusual names, locations, and other proper nouns Technical or industry-specific terms Out of Vocabulary / not in the dictionary HomonymsSmart phone Tablet Laptop Desktop Microphone Array Headset Built-in Hand held Camera Specialist /Smart Computer MobileDirect audio recording Synthetic speech recording Noise-network distorted speech Connectivity Live / Real-Time Recorded  
Table 1. Additional practical metrics to support the evaluation of ASR outcomes

When it comes to pronunciation or typical morphosyntactic differences in the way a language is used, developers may be able to pre-empt and automate corrections for consistent errors. An example includes articulation errors that are typical for those speaking English as a foreign language such as the omission of “th”, “v” and “rl” sounds that do not appear in some Chinese dialects. 

girl holding a mobile with cide indicating interaction.

Age and gender biases could also be improved using semi-automated annotation systems, but speaker style remains an issue that is hard to change when there is direct ‘human-machine interaction’ rather than someone reading text.

Moreover, there still remains the manual process of checking for metrics, such as those that examine the way technology is used. This type of problem can be judged visually if the camera catches the interactions and in an auditory manner, such as walking away from the microphone or turning ones back to the camera etc. AI Video content analysis is moving apace and these techniques could help us in time!

Ultimately the training data is the main issue but automated bias mitigation techniques are being explored by researchers and the outcomes look promising and there also needs to be some swift designing of a more sophisticated and adaptable ASR performance metric evaluator to automate the process of reviewing output!